Pasar al contenido principal

METODOS FORMALES ESCALABLES PARA APLICACIONES REALES

PID2021-122830OB-C42

•
Nombre agencia financiadora Agencia Estatal de Investigación
Acrónimo agencia financiadora AEI
Programa Programa Estatal para Impulsar la Investigación Científico-Técnica y su Transferencia
Subprograma Subprograma Estatal de Generación de Conocimiento
Convocatoria Proyectos de I+D+i (Generación de Conocimiento y Retos Investigación)
Año convocatoria 2021
Plan Estatal 2021-2023 Plan Estatal de Investigación Científica y Técnica y de Innovación
Centro beneficiario UNIVERSITAT POLITÈCNICA DE VALÈNCIA

Resultados relacionados

Resultados totales (Incluyendo duplicados): 33
Encontrada(s) 1 página(s)

RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

General scales unlock AI evaluation with explanatory and predictive power

  • Zhou, Lexin
  • Pacchiardi, Lorenzo
  • Collins, Katherine M.
  • Zhang, Seraphina
  • Zhao, Qinlin
  • Huang, Yitian
  • Sun, Luning
  • Prunty, Jonathan E.
  • Li, Zongqian
  • Sanchez-Garcia, Pablo
  • Casares, Pablo A. M.
  • Zu, Jiyun
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Moros-Daval, Yael
  • Jiang-Chen, Ke-Xin
  • Burden, John
  • Mehrbakhsh, Behzad|||0000-0001-9017-989X
  • Stillwell, David
  • Cebrian, Manuel
  • Wang, Jindong
  • Henderson, Peter
  • Wu, Sherry Tongshuang
  • Kyllonen, Patrick C.
  • Cheke, Lucy
  • Xie, Xing
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Ensuring safe and effective use of artificial intelligence (AI) requires understanding and anticipating its performance on new tasks, from advanced scientific challenges to transformed workplace activities1, 2-3. So far, benchmarking has guided progress in AI but has offered limited explanatory and predictive power for general-purpose AI systems4, 5, 6, 7-8, attributed to limited transferability across specific tasks9, 10-11. Here we introduce general scales for AI evaluation that elicit demand profiles explaining what capabilities common AI benchmarks truly measure, extract ability profiles quantifying the general strengths and limits of AI systems and robustly predict AI performance for new task instances. Our fully automated methodology builds on 18 rubrics, capturing a broad range of cognitive and intellectual demands, which place different task instances on the same general scales, illustrated on 15 large language models (LLMs) and 63 tasks. Both the demand and the ability profiles on these scales bring new insights such as construct validity through benchmark sensitivity and specificity and explain conflicting claims about whether AI has reasoning capabilities. Ultimately, high predictive power at the instance level becomes possible using the general scales, providing superior estimates over strong black-box baseline predictors, especially in out-of-distribution settings (new tasks and benchmarks). The scales, rubrics, battery, techniques and results presented here constitute a solid foundation for a science of AI evaluation, underpinning the reliable deployment of AI in the years ahead., We thank the POLARIS Lab and the Computational Cognitive Science Lab at Princeton University, the European AI Office, the UK AISI, the Future of Life Institute, the OECD Future of Skills team, S. Chang, M. Zilka, J. Lian, R. Burnell, W. Schellaert, A. Gomez, M. Tesic, C. Li and P. Romero for their valuable help and feedback at certain stages of the project. We thank OpenAI for granting us research access to several LLMs to conduct the experiments in this paper and DeepSeek and Meta for giving us access to the weights of their models. We acknowledge support from the following institutions: Microsoft Accelerate Foundation Models Research (AFMR) grant programme, Long-Term Future Scholarship financed by Coefficient Giving (formerly Open Philanthropy) and the Spanish Government's Knowledge Generation Projects (PID2023-150271NB-C21). This work was also supported by CIPROM/2022/6 (FASSLOW), IDIFEDER/2021/05 (CLUSTERIA), CIACIF/2023/276, financed by Generalitat Valenciana, the EC H2020-EU grant agreement no. 952215 (TAILOR), Spanish grants PID2021-122830OB-C42 (SFERA), PID2023-150271NB-C21 and PID2024-162030OB-100 (ROBIN) financed by MCIN/AEI/10.13039/501100011033 and 'ERDF A way of making Europe', Catedra ENIA-UPV in Sustainable AI Development, TSI-100930-2023-9, INCIBE's Chair financed by the EU's NextGenerationEU, EUR2024-153548 (PREDAIT) 'Towards Predictable AI' from 'Spanish Europe Excelencia' 2024, Spanish National Research Council (CSIC) Special Intramural Projects programme and the Cambridge Trust. M.C. declares support from Google.org through the Silicon Valley Community Foundation by means of a grant to Fundacion General CSIC. The research of J.H.-O. is supported by OpenAI's grant to the 'AI Progress through the Lens of Predictable AI Ecosystems' programme, which is based at the Leverhulme Centre for the Future of Intelligence at the University of Cambridge.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

An Efficient Canonical Narrowing Implementation with Irreducibility and SMT Constraints for Generic Symbolic Protocol Analysis

  • López-Rueda, Raúl|||0000-0002-2550-8041
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Sapiña-Sanchis, Julia|||0000-0003-2994-6986
[EN] Narrowing and unification are very useful tools for symbolic analysis of rewrite theories, and thus for any model that can be specified in that way. A very clear example of their application is the field of formal cryptographic protocol analysis, which is why narrowing and unification are used in tools such as Maude-NPA, Tamarin and Akiss. In this work we present the implementation of a canonical narrowing algorithm, which improves the standard narrowing algorithm, extended to be able to process rewrite theories with conditional rules. The conditions of the rules will contain SMT constraints, which will be carried throughout the execution of the algorithm to determine if the solutions have associated satisfiable or unsatisfiable constraints, and in the latter case, discard them., This work has been partially supported by the EC H2020-EU grant agreement No. 952215 (TAILOR) , by the grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe, by the grant CIPROM/2022/6 funded by Generalitat Valenciana, and by the grant PCI2020-120708-2 funded by MICIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR. Raul Lopez has been supported by the grant FPI-2022- S2-49802 funded by Universitat Politecnica de Valencia.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Rethink reporting of evaluation results in AI

  • Burnell, Ryan
  • Schellaert, Wout Willy M.
  • Burden, John
  • Ullman, Tomer D.
  • Tenenbaum, Joshua B.
  • Rutar, Danaja
  • Cheke, Lucy G.
  • Sohl-Dickstein, Jascha
  • Mitchell, Melanie
  • Kiela, Douwe
  • Shanahan, Murray
  • Voorhees, Ellen M.
  • Cohn, Anthony
  • Leibo, Joel Z.
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Artificial intelligence (AI) systems have begun to be deployed in high-stakes contexts, including autonomous driving and medical diagnosis. In contexts such as these, the consequences of system failures can be devastating. It is therefore vital that researchers and policymakers have a full understanding of the capabilities and weaknesses of AI systems so that they can make informed decisions about where these systems are safe to use and how they might be improved., We are grateful to J. Pineau and S. Russell for their feedback on the manuscript and to M. Brundage and L. Ahmad for helpful discussions about the need for instance-by-instance results for foundation models. Funding support was provided by the Future of Life Institute under grant RFP2-152 (J.H.O. and J.B.); US Defense Advanced Research Projects Agency HR00112120007 (RECoG-AI) (J.H.O., L.C., R.B., J.B., and D.K.); grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe (J.H.O., F.M.P., and W.S.); INNEST/2021/317 [Project cofunded by the European Union with the Programa Operativo del Fondo Europeo de Desarrollo Regional (FEDER) de la Comunitat Valenciana 2014-2020 ] (J.H.O. and W.S.); and UPV (Vicerrectorado de Investigación) grant PAI-10-21 (J.H.O. and W.S.).
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Unveiling the robustness of machine learning families

  • Fabra-Boluda, Raúl
  • Ferri, C.
  • Ramírez Quintana, María José|||0000-0002-0559-3568
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
[EN] The evaluation of machine learning systems has typically been limited to performance measures on clean and curated datasets, which may not accurately reflect their robustness in real-world situations where data distribution can vary from learning to deployment, and where truthfully predict some instances could be more difficult than others. Therefore, a key aspect in understanding robustness is instance difficulty, which refers to the level of unexpectedness of system failure on a specific instance. We present a framework that evaluates the robustness of different ML models using item response theory-based estimates of instance difficulty for supervised tasks. This framework evaluates performance deviations by applying perturbation methods that simulate noise and variability in deployment conditions. Our findings result in the development of a comprehensive taxonomy of ML techniques, based on both the robustness of the models and the difficulty of the instances, providing a deeper understanding of the strengths and limitations of specific families of ML models. This study is a significant step towards exposing vulnerabilities of particular families of ML models., This work was funded by the Norwegian Research Council Grant 329745 Machine Teaching for Explainable AI, the MIT-Spain-INDITEX Sustainability Seed Fund under Project COST-OMIZE, CIPROM/2022/6 (FASSLOW) funded by Generalitat Valenciana, the EC H2020-EU Grant Agreement No. 952215 (TAILOR), and Spanish Grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and 'ERDF A way of making Europe'. RFB is supported by predoctoral Grant PRE2019-090892.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Detection of microfibres in wastewater sludge with deep learning

  • Martí-Pérez, Félix
  • Domínguez-Rodríguez, Ana|||0009-0003-6175-2361
  • C. Monserrat|||0000-0003-1790-8085
  • Ferri Ramírez, César|||0000-0002-8975-1120
  • Lujan Facundo, Maria Jose|||0000-0001-6871-0584
  • Ferrer-Polonio, Eva|||0000-0001-6412-3825
  • Bes-Piá, M.A.|||0000-0001-8005-7450
  • Mendoza Roca, José Antonio
[EN] The proliferation of microplastics, especially microfibres (MFi), represents a significant environmental concern due to their persistence and potential health risks. In particular, wastewater treatment plants are among the primary contributors to the release of MFi into ecosystems, making it crucial to detect and quantify these pollutants, mainly in sewage sludge. Effective monitoring not only helps assess the extent of pollution but also identifies key sources. Traditional detection methods are labour-intensive and lack the scalability necessary for effective monitoring. This study presents a novel approach utilising advanced deep learning techniques to enhance the detection of MFi in sewage sludge samples using two different filtration support (fibreglass and cellulose acetate). By leveraging convolutional neural networks (CNNs), we developed a robust system for accurately identifying and localising microfibres. Our deep learning framework, implemented using Mask R-CNN architecture, demonstrates superior performance in detecting MFi, achieving a mean average precision (mAP) of 72% for the glass dataset and 68% for the cellulose acetate dataset. This approach significantly reduces manual effort and processing time. However, further improvements are still possible, as our model shows weaker performance for smaller fibres and those that resemble fibrous morphology (particularly on cellulose acetate filters). Future work could address these issues by expanding the presented datasets to address the identified shortcomings., This work was supported by the project CIPROM/2022/6 (FASSLOW) funded by Generalitat Valenciana, the EC H2020-EU grant agreement No.952215 (TAILOR), and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ``ERDF A way of making Europe''. Authors thank Grant PID2021-127468OBI00 funded by MICIU/AEI/10.13039/501100011033 and by ERDF/EU. The authors also thank Business Chair FACSA-FOVASA from Universitat Politècnica de València.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Analysing the Predictability of Language Model Performance

  • Schellaert, Wout Willy M.
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Can a language model predict for which questions another language model will answer successfully? We investigate the extent to which performance prediction is possible and dissect various factors that influence it. Our experimental setting fine-tunes DeBERTa models, which we call assessors, on the evaluation results of generative language models with up to 128 billion parameters, which we refer to as subject systems. Our analysis spans more than 100 tasks from BIG-bench. We find that the assessors can match and even exceed the subjects' confidence in both refinement and calibration, anticipating failures at near perfect levels for some tasks. We also find that for performance prediction it can be beneficial to learn from the scores on multiple tasks or to learn from the scores of multiple subjects, but both depend on the task at hand. Lastly, we find that large and small subject systems are equally predictable, showing promise for the scalability of the predictability problem., This work was funded by valgrAI, the Future of Life Institute (FLI) under grant RFP2-152, the EU (FEDER) and Spanish grant RTI2018-094403-B-C32 funded by MCIN/AEI/10.13039/501100011033 and by CIPROM/2022/6 funded by Generalitat Valenciana, FISCALTICS (I+D+i PID2022-140110OA-I00) granted by MICIU/AEI/10.13039/501100011033, the Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and 'ERDF A way of making Europe, as well as the European Union under agreement INNEST/2021/317 (Neurocalcat) and by the Vic. Inv. of the Universitat Politecnica de Valencia under DOCEMPR21, DOCEMPR22, and DOCEMPR23'.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Larger and more instructable language models become less reliable

  • Zhou, Lexin
  • Schellaert, Wout Willy M.
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Moros-Daval, Yael
  • Ferri Ramírez, César|||0000-0002-8975-1120
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] The prevailing methods to make large language models more powerful and amenable have been based on continuous scaling up (that is, increasing their size, data volume and computational resources1 ) and bespoke shaping up (including post-fltering2,3 , fne tuning or use of human feedback4,5 ). However, larger and more instructable large language models may have become less reliable. By studying the relationship between difculty concordance, task avoidance and prompting stability of several language model families, here we show that easy instances for human participants are also easy for the models, but scaled-up, shaped-up models do not secure areas of low difculty in which either the model does not err or human supervision can spot the errors. We also fnd that early models often avoid user questions but scaled-up, shaped-up models tend to give an apparently sensible yet wrong answer much more often, including errors on difcult questions that human supervisors frequently overlook. Moreover, we observe that stability to diferent natural phrasings of the same question is improved by scaling-up and shaping-up interventions, but pockets of variability persist across difculty levels. These fndings highlight the need for a fundamental shift in the design and development of general-purpose artifcial intelligence, particularly in high-stakes areas for which a predictable distribution of errors is paramount., We acknowledge support from the following institutions: Long-Term Future Scholarship funded by Open Philanthropy; Grant for Master Studies funded by ValgrAI Valencian Graduate School and Research Network for Artificial Intelligence and Generalitat Valenciana; FISCALTICS (I+D+i PID2022-140110OA-I00), funded by MICIU/AEI/10.13039/501100011033 and by ERDF, EU; CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana; the EC H2020-EU grant agreement no. 952215 (TAILOR); US DARPA HR00112120007 (RECoG-AI) and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe ; and the Vic. Inv. of the Universitat Politècnica de Valencia under DOCEMPR21, DOCEMPR22 and DOCEMPR23.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

When Redundancy Matters: Machine Teaching of Representations

  • Ferri Ramírez, César|||0000-0002-8975-1120
  • Hernández-Orallo, José|||0000-0001-9746-7632
  • Garigliotti, Dario
  • Havardstun, Brigt
  • Telle, Jan Arne
[EN] In traditional machine teaching, a teacher needs to teach a concept to a learner by means of a finite set of examples, the witness set. But concepts can have many equivalent representations. This redundancy strongly affects the search space, to the extent that teacher and learner may not be able to easily determine the equivalence class of each representation. In this common situation, instead of teaching concepts, we explore the idea of teaching representations. We work with several teaching schemas that exploit representation and witness size (Eager, Greedy and Optimal) and analyze the gains in teaching effectiveness, both theoretically, and also experimentally for languages where redundancy can vary (DNF expressions and Turing-complete P3 programs). Our theoretical and experimental results indicate that there are various types of redundancy, related, e.g,. to the spread of the redundant representations, handled better by the new Greedy schema introduced here than by the Eager schema. For P3 programs witness sets found by Greedy are usually smaller than the programs they identify, corroborating previous results that conveying information efficiently is a leitmotif of machine teaching., Open access funding provided by University of Bergen (incl Haukeland University Hospital). This work was funded by the Norwegian Research Council grant 329745 Machine Teaching for Explainable AI, CIPROM/2022/6 (FASSLOW) funded by Generalitat Valenciana, the EC H2020-EU grant agreement No. 952215 (TAILOR), and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN / AEI / 10.13039/501100011033 and ERDF A way of making Europe.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Contribución de congreso (ConferenceOutput)

Protocol Dialects as Formal Patterns

  • Galán, D.
  • Meadows, C.
  • Meseguer, J.
  • García-Valero, Víctor|||0000-0003-0681-1130
  • Escobar Román, Santiago|||0000-0002-3550-4781
[EN] A protocol dialect is a lightweight method to obfuscate the communication exchanges between legitimate protocol users to make it hard for malicious users to interact with legitimate ones. So far, dialects have been based on a single obfuscation transformation, which we call a lingo. In this work dialects are generalized to become protocol and lingo generic. In this way they can be composed with other dialects for greater security. We present a formal semantics of dialects as formal patterns, i.e., as protocol transformations formalized in rewriting logic. We also present several attacker models and explain how our generalized notion of dialect can be used to harden dialects against such attackers., S. Escobar and V. Garcia have been partially supported by the grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe, by the grant CIPROM/2022/6 funded by Generalitat Valenciana, and by the grant PCI2020-120708-2 funded by MICIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR. D. Galan, V. Garcia, C. Meadows, and J. Meseguer have been partially supported by ONR Code 311.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Team formation through an assessor: choosing MARL agents in pursuit evasion games

  • Zhao, Yue
  • Ju, Lushan
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Team formation in multi-agent systems usually assumes the capabilities of each team member are known, and the best formation can be derived from that information. As AI agents become more sophisticated, this characterisation is becoming more elusive and less predictive about the performance of a team in cooperative or competitive situations. In this paper, we introduce a general and flexible way of anticipating the outcome of a game for any lineups (the agents, sociality regimes and any other hyperparameters for the team). To this purpose, we simply train an assessor using an appropriate team representation and standard machine learning techniques. We illustrate how we can interrogate the assessor to find the best formations in a pursuit-evasion game for several scenarios: offline team formation, where teams have to be decided before the game and not changed afterwards, and online team formation, where teams can see the lineups of the other teams and can be changed at any time., This work was funded by CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana, the EC H2020-EU grant agreement No. 952215 (TAILOR), and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Local confluence of conditional and generalized term rewriting systems

  • Lucas Alba, Salvador|||0000-0001-9923-2108
[EN] Reduction-based systems are used as a basis for the implementation of programming languages, automated reasoning systems, mathematical analysis tools, etc. In such inherently non deterministic systems, guaranteeing that diverging steps can be eventually rejoined is crucial for a faithful use in most applications. This property of reduction systems is called local confluence. In a landmark 1980 paper, Gerard Huet characterized local confluence of a Term Rewriting System as the joinability of all its critical pairs. In this paper, we characterize local confluence of Conditional Term Rewriting Systems, where reduction steps may depend on the satisfaction of specific conditions in rules: a conditional term rewriting system is locally confluent if and only if (i) all its conditional critical pairs and (ii) all its conditional variable pairs (which we introduce in this paper) are joinable. Furthermore, the logic-based approach we follow here is well-suited to analyze local confluence of more general reduction-based systems. We exemplify this by (i) including (context-sensitive) replacement restrictions in the arguments of function symbols, and (ii) allowing for more general conditions in rules. The obtained systems are called Generalized Term Rewriting Systems. A characterization of local confluence is also given for them., Partially supported by grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and by "ERDF A way of making Europe" and by the grant CIPROM/2022/6 funded by Generalitat Valenciana.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Parallel Maude-NPA for Cryptographic Protocol Analysis

  • Do, Canh Minh
  • Riesco, Adrian
  • Ogata, Kazuhiro
  • Escobar Román, Santiago|||0000-0002-3550-4781
[EN] Maude-NPA is a formal verification tool for analyzing cryptographic protocols in the Dolev-Yao strand space model modulo an equational theory defining the cryptographic primitives. It starts from an attack state to find counterexamples or conclude that the attack concerned cannot be conducted by performing a backward narrowing reachability analysis. Although Maude-NPA is a powerful analyzer, its running performance can be improved by taking advantage of parallel and/or distributed computing when dealing with complex protocols whose state space is huge. This paper describes a parallel version of Maude-NPA in which the backward narrowing and the transition subsumption at each layer in Maude-NPA are conducted in parallel. A tool supporting the parallel version has been implemented in Maude with a master-worker model using meta-interpreters. We report on some experiments of various kinds of protocols that demonstrate that the tool can increase the running performance of Maude-NPA by 52% on average for complex case studies in which the number of states located at each layer is considerably large., This workwas supported in part by the JST SICORP under Grant JPMJSC20C2, Japan,in part by JSPS KAKENHI under Grant JP23K28060, Grant JP23K19959,Grant JP24K20757, Grant JP24KK0185, in part by the ProCode10 under Grant PID2023-149943OB-I00, in part by MICIN under Grant PID2021-122830OB-C42, in part by MCIN/AEI/10.13039/501100011033, in part by ERDF A way ofmaking Europe under Grant CIPROM/2022/6, in part by Generalitat Valenciana,and in part by INCIBE's Chair funded by the EU-NextGenerationEU throughthe Spanish government's Plan de Recuperacion, Transformacion y Resiliencia.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Formal Analysis of Post-Quantum Hybrid Key Exchange SSH Transport Layer Protocol

  • Tran, Duong Dinh
  • Ogata, Kazuhiro
  • Akleylek, Sedat
  • Otmani, Ayoub
  • Escobar Román, Santiago|||0000-0002-3550-4781
[EN] Facing the quantum attack threat, a quantum-resistant version of the SSH Transport Layer protocol has been proposed and been standardized by an IETF working group. This standardization process has been motivated by the fact that if practical quantum computers become available, classical key exchange algorithms used today will no longer be safe because their security can be efficiently broken by Shor's algorithm running on a quantum computer. In this paper, we construct a symbolic model of the proposed protocol, specify it in the specification language CafeOBJ, and conduct a formal analysis of the claimed security properties with the employment of a formal method tool called Invariant Proof Score Generator (IPSG). Three security properties are formally verified with respect to an unbounded number of protocol participants and protocol executions by employing IPSG to produce their formal proofs, the so- called proof scores in CafeOBJ. With another property, namely authentication, we find a counterexample against it, and then we propose to slightly revise the protocol. With the improved version, we formally verify that the authentication property is enjoyed, while the three properties remain secure. To model the presence of malicious participants, we extend the Dolev-Yao intruder, which is the standard intruder model used in security protocol analysis/verification, because the availability assumption of large-scale quantum computers gives the attacker some new capabilities. Under our threat model, the intruder is given capabilities of fully controlling the network as the Dolev-Yao intruder, and additionally breaking ECDH's security as well as compromising secrets., The work of Duong Dinh Tran and Kazuhiro Ogata was supported by the JST SICORP, Japan, under Grant JPMJSC20C2. The work of Santiago Escobar was supported in part by the MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe under Grant PID2021-122830OB-C42, in part by the Generalitat Valenciana under Grant CIPROM/2022/6, in part by the MICIN/AEI/10.13039/501100011033 under Grant PCI2020-120708-2, and in part by the European Union NextGenerationEU/PRTR. The work of Sedat Akleylek was supported in part by TUBITAK under Grant 121R006. The work of Ayoub Otmani was supported in part by the FAVPQC Project funded by CNRS, and in part by the Agence Nationale de la Recherche (ANR) within France 2030 Program under Grant ANR-22-PETQ-0008 PQ-TLS.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Trends in AI inference energy consumption: Beyond the performance-vs-parameter laws of deep learning

  • Desislavov, Radosvet
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] The progress of some AI paradigms such as deep learning is said to be linked to an exponential growth in the number of parameters. There are many studies corroborating these trends, but does this translate into an exponential increase in energy consumption? In order to answer this question we focus on inference costs rather than training costs, as the former account for most of the computing effort, solely because of the multiplicative factors. Also, apart from algorithmic innovations, we account for more specific and powerful hardware (leading to higher FLOPS) that is usually accompanied with important energy efficiency optimisations. We also move the focus from the first implementation of a breakthrough paper towards the consolidated version of the techniques one or two year later. Under this distinctive and comprehensive perspective, we analyse relevant models in the areas of computer vision and natural language processing: for a sustained increase in performance we see a much softer growth in energy consumption than previously anticipated. The only caveat is, yet again, the multiplicative factor, as future AI increases penetration and becomes more pervasive., We thank the reviewers for their insightful remarks, which have helped improve the paper significantly. This work has been partially supported by the MIT-Spain-INDITEX Sustainability Seed Fund under project COST-OMIZE, the grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe", Generalitat Valenciana under INNEST/2021/317 and PROMETEO/2019/098, EU's Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR) , the Future of Life Institute, FLI, under grant RFP2-152, US DARPA HR00112120007 (RECoG-AI).
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Contamination Budget: Trade-offs between Breadth, Depth and Difficulty

  • Mehrbakhsh, Behzad|||0000-0001-9017-989X
  • Hernández-Orallo, José|||0000-0001-9746-7632
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
Contamination in large language models (LLMs), and machine learning more broadly, refers to the inclusion of equal –or very similar– examples in both training and test sets. This phenomenon usually translates into better test performance. Here we explore when this contamination is performed intentionally, for purposes that can be malicious (e.g., get better scores in evaluations) or benevolent (e.g., fix some mistakes). These interventions, usually in the form of fine-tuning memorisations, come with a budget in the size of the fine-tuning dataset. Several trade-offs appear between the breadth of the intervention (how many examples to be memorised), its depth (how many repetitions of each example) and the difficulty of the examples. By studying several LLMs and datasets, we observe some monotonic behaviour (more difficult items require more depth to be ‘fixed’) but also some non-monotonic phenomena (very high depth levels have negative effects on non-contaminated examples). This suggests that trade-offs should be found not only in terms of the budget but also according to model specifics, the task and the item difficulty at hand., We acknowledge support from: Grant for Master Studies funded by ValgrAI – Valencian Graduate School and Research Network for Artificial Intelligence and Generalitat Valenciana; granted by MICIU/AEI/10.13039/ 501100011033 and by ERDF, EU; CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana; and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and “ERDF A way of making Europe”.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Contribución de congreso (ConferenceOutput)

Distilling the Effects of Language Model Contamination

  • Mehrbakhsh, Behzad|||0000-0001-9017-989X
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] The proportion of AI-generated content permeating the well of knowledge is increasing significantly. Large language models (LLMs) contribute to that contamination but they also suffer from it. However, it is yet to be clarified the effect of different sources of error, be it human-generated or LLM-generated. Controlling for the percentage of error, we explore the impact on LLM fine-tuning when errors come from humans, from other language models or are generated randomly using an aleatoric or epistemic source. In this paper, we compare these different types of error for in-distribution and out-of-distribution experimental settings. By analysing the levels of errors and their distribution, we find a nuanced view: while in-distribution human-generated noise seems more benign than the LLM-generated counterpart, in the out-of-distribution case the model-generated noise may not be necessarily worse., We thank the anonymous reviewers for their comments and Dr. Peter Clark for providing us with the extended version of the OpenBookQA dataset. This work was funded by valgrAI, the Future of Life Institute, FLI, under grant RFP2-152, the EU (FEDER) and Spanish grant RTI2018-094403-B-C32 funded by MCIN/AEI/10.13039/501100011033 and by CIPROM/2022/6 funded by Generalitat Valenciana, EU s Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR), and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe"
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

A general supply-inspect cost framework to regulate the reliability-usability trade-offs for few-shot inference

  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Ramírez Quintana, María José|||0000-0002-0559-3568
  • Hernández-Orallo, José|||0000-0001-9746-7632
  • Jaimovitch-López, Gonzalo Eduardo
  • Ferri, C.
[EN] Language models and other recent machine learning paradigms blur the distinction between generative and discriminative tasks, in a continuum that is regulated by the degree of pre- and post-supervision that is required from users, as well as the tolerated level of error. In few-shot inference, we need to find a trade-off between the number and cost of the solved examples that have to be supplied, those that have to be inspected (some of them accurate but others needing correction) and those that are wrong but pass undetected. In this paper, we define a new Supply-Inspect Cost Framework, associated graphical representations and comprehensive metrics that consider all these elements. To optimise few-shot inference under specific operating conditions, we introduce novel algorithms that go beyond the concept of rejection rules in both static and dynamic contexts. We illustrate the effectiveness of all these elements for a transformative domain, data wrangling, for which language models can have a huge impact if we are able to properly regulate the reliability-usability trade-off, as we do in this paper., The funding has been received from ValgrAI - Valencian Graduate School and Research Network for Artificial Intelligence; the Norwegian Research Council with Grant no. 329745 (Machine Teaching for Explainable AI); Generalitat Valenciana with Grant nos. CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA); the European Commission under H2020-EU with Grant no. 952215 (TAILOR); US DARPA with Grant no. HR00112120007 (RECoG-AI); the Future of Life Institute with Grant no. RFP2-152; the Spanish Ministry of Science and Innovation (MCIN/AEI/10.13039/ 501100011033) with Grant no. PID2021-122830OB-C42 (SFERA) and ERDF A way of making Europe ; and the Spanish Ministry of Universities with Grant no. PID2022-140110OA-I00 (FISCALTICS) funded by MICIU/AEI/10.13039/501100011033 and by ERDF, EU.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Training Data Scientists Through Project-Based Learning

  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] The concepts of innovation, creativity, problem solving, effective communication, autonomy and critical thinking are at the core of becoming a good data scientist. Adapting to new technological resources and tools is also an important skill, which also builds on the curious and inquisitive nature associated with data science, and is fuelled by rapidly changing data science ecosystems in industry. In this regard, Project-based learning (PBL) has clear benefits for engaging students in data science courses. However, the exploratory character of data science projects, which do not start with a clear specification of what to do, but some data to analyse, pose some challenges to the application of PBL. Our aim is to improve students' data science learning experiences and outcomes through the use of PBL. In this paper, we share our experiences with PBL and present an assessment rubric that focuses on value, innovation and narrative, which can be used as a scaffolding structure for data science courses. Our analysis of a PBL data science course at MSc level, together with data from student surveys, shows how the methodology and rubric align well with the exploratory nature of data science and the proactive, curious, and inquisitive skills required of data scientists., Thiswork was supported in part by valgrAI, the Norwegian Research Council Grant 329745 Machine Teaching for Explainable AI, the Future of LifeInstitute (FLI), under Grant RFP2-152, the EU (FEDER), and Spanish Grant RTI2018-094403-B-C32 through MCIN/AEI/10.13039/501100011033 and in part by CIPROM/2022/6 through Generalitat Valenciana, EU's Horizon 2020 Research and Innovation Programme under Grant 952215 (TAILOR), U.S. DARPA HR00112120007 (RECoG-AI), and Spanish Grant PID2021-122830OB-C42 (SFERA) through MCIN/AEI/10.13039/501100011033 and "ERDF a Way of Making Europe."
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Formalization and analysis of the post-quantum signature scheme FALCON with Maude

  • García-Valero, Víctor|||0000-0003-0681-1130
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Ogata, Kazuhiro
[EN] Digital signatures ensure the authenticity and integrity of digital assets, vital properties for any secure communication. The National Institute of Standards and Technologies launched the Post-Quantum Cryptography project to standardise new algorithms and protocols that are secure against quantum attackers. The post-quantum signature scheme FALCON was one of the finalists. We present a continuation of the first steps towards the formal specification and analysis, in the high-performance language Maude, of signature schemes. We have adapted and improved a previous framework, originally aimed to formally specify and analyse post-quantum key encapsulation mechanisms. As a use case of the new framework, we specify an executable symbolic model of FALCON. On the symbolic model, we verify termination and fairness using LTL formulas with Maude's model checker. Furthermore, authentication, integrity and non-repudiation are analysed through invariant analysis. Integrity and non-repudiation hold, meanwhile, authentication does not hold in our symbolic model., star Santiago Escobar and Victor Garcia have been partially supported by the grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe, by the grant CIPROM/2022/6 funded by Generalitat Valenciana, by INCIBE's Chair funded by the EU-NextGenerationEU through the Spanish government's Plan de Recuperacion, Transformacion y Resiliencia, and by the grant PCI2020-120708-2 funded by MICIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR. Kazuhiro Ogata has been partially supported by JSPS KAKENHI Grant Number 24KK0185.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Can language models automate data wrangling?

  • Jaimovitch-López, Gonzalo
  • Ferri Ramírez, César|||0000-0002-8975-1120
  • Hernández-Orallo, José|||0000-0001-9746-7632
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Ramírez Quintana, María José|||0000-0002-0559-3568
[EN] The automation of data science and other data manipulation processes depend on the integration and formatting of 'messy' data. Data wrangling is an umbrella term for these tedious and time-consuming tasks. Tasks such as transforming dates, units or names expressed in different formats have been challenging for machine learning because (1) users expect to solve them with short cues or few examples, and (2) the problems depend heavily on domain knowledge. Interestingly, large language models today (1) can infer from very few examples or even a short clue in natural language, and (2) can integrate vast amounts of domain knowledge. It is then an important research question to analyse whether language models are a promising approach for data wrangling, especially as their capabilities continue growing. In this paper we apply different variants of the language model Generative Pre-trained Transformer (GPT) to five batteries covering a wide range of data wrangling problems. We compare the effect of prompts and few-shot regimes on their results and how they compare with specialised data wrangling systems and other tools. Our major finding is that they appear as a powerful tool for a wide range of data wrangling tasks. We provide some guidelines about how they can be integrated into data processing pipelines, provided the users can take advantage of their flexibility and the diversity of tasks to be addressed. However, reliability is still an important issue to overcome., Open Access funding provided thanks to the CRUE-CSIC agreement with Springer Nature. This work was funded by the Future of Life Institute, FLI, under grant RFP2-152, the MIT-Spain - INDITEX Sustainability Seed Fund under project COST-OMIZE, the EU (FEDER) and Spanish MINECO under RTI2018-094403-B-C32 and PID2021-122830OB-C42, Generalitat Valenciana under PROMETEO/2019/098 and INNEST/2021/317, EU's Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR) and US DARPA HR00112120007 ReCOG-AI. AcknowledgementsWe thank Lidia Contreras for her help with the Data Wrangling Dataset Repository. We thank the anonymous reviewers from ECMLPKDD Workshop on Automating Data Science (ADS2021) and the anonymous reviewers of this special issue for their comments.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Modeling and verification of the post-quantum key encapsulation mechanism KYBER using Maude

  • García-Valero, Víctor|||0000-0003-0681-1130
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Ogata, Kazuhiro
[EN] Communication and information technologies shape the world¿s systems of today, and those systems shape the society we live in. The security of those systems relay on mathematical problems hard to solve for classical computers, that is, the available current computers. Recent advances in quantum computing threat the security of our systems and the communications we use. In order to face the threat, multiple solutions and protocols have been proposed. Kyber is one of them, specifically it is a key encapsulation mechanism that bases its security in the learning with errors problem over module lattices. This master¿s thesis focuses on the analysis of Kyber to check its security under Dolev-Yao adversary assumptions. For that matter, we first learn about the current state of the solutions proposed against the threat quantum adversaries suppose and study how Kyber works. We then construct in the system-specification language Maude, a symbolic model to represent the behaviour of Kyber in a network. On this model we conduct reachability analysis with the search command and find that a Man-In-The-Middle attack is present. Then we use the Maude LTL logical model checker to extend the analysis of the system with proving if a liveness and a security property hold.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Follow the leader: a deep reinforcement learning framework for safe and efficient autonomous car-following

  • Marin, Andres L.
  • Makridis, Michail A.
  • Tansini, Alessandro
  • Pulvirenti, Luca
  • Corujo, Jaime Suarez
  • Komnos, Dimitrios
  • Fontaras, Georgios
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Ramírez Quintana, María José|||0000-0002-0559-3568
  • C. Monserrat|||0000-0003-1790-8085
[EN] Autonomous and semi-autonomous driving technologies present a promising path to achieving safer and more efficient transportation, potentially reducing energy use and emissions. However, these advantages might not be realized without carefully targeted optimization and ad-hoc designed dataset to evaluate the efficacy of the optimization performance. Through the analysis of both driver actions and vehicle dynamics using a real-world dataset, we present an innovative reward function based on the Deep Deterministic Policy Gradient (DDPG) method. The reward function aims to minimize energy consumption and the increased travel time while keeping acceleration within physical limits. The approach results in an average of 7% energy savings in various vehicle powertrains, including Internal Combustion Engine Vehicles (ICEV), Plug-in Hybrid Electric Vehicles (PHEV), and Battery Electric Vehicles (BEV), while keeping acceleration within realistic ranges, maintaining safe following distances, and keeping trip time increases below 1%. The dataset can be accessed at https://data.jrc.ec.europa.eu/collection/id-00437., We acknowledge support from: MICIU/AEI/10.13039/501100011033 and by ERDF, EU; CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana; and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe," Catedra ENIA-UPV in Sustainable AI Development, TSI-100930-2023-9, and INCIBE's Chair funded by the EU-NextGenerationEU through the Spanish government's Plan de Recuperacion, Transformacion y Resiliencia. The authors would also like to acknowledge the support of the Joint Research Centre.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

A Partial Evaluation Methodology for Optimizing Rewrite Theories Incrementally

  • Alpuente Frasnedo, María|||0000-0002-9268-1178
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Sapiña-Sanchis, Julia|||0000-0003-2994-6986
  • Ballis, Demis
  • Galán-Pascual, Daniel
[EN] Partial evaluation (PE) is a branch of computer science that achieves code optimization via specialization. This article describes a PE methodology for optimizing rewrite theories that encode concurrent as well as nondeterministic systems by means of the Maude language. The main advantages of the proposed methodology can be summarized as follows: An automatic program optimization technique for rewrite theories featuring several PE criteria that support the specialization of a broad class of rewrite theories. An incremental partial evaluation modality that allows the key specialization components to be encapsulated at the desired granularity level to facilitate progressive refinements of the specialization. All executability theory requirements are preserved by the PE transformation. Also the transformation ensures the semantic equivalence between the original rewrite theory and the specialized theory under rather mild conditions., This work was partially supported by TAILOR, a project funded by the EU Horizon 2020 research and innovation programme under GA No 952215, grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and by "ERDF A way of making Europe", and by the Generalitat Valenciana under grant PROMETEO/2019/098.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Modelling and verification of post-quantum key encapsulation mechanisms using Maude

  • García-Valero, Víctor|||0000-0003-0681-1130
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Ogata, Kazuhiro
  • Akleylek, Sedat
  • Otmani, Ayoub
[EN] Communication and information technologies shape the world's systems of today, and those systems shape our society. The security of those systems relies on mathematical problems that are hard to solve for classical computers, that is, the available current computers. Recent advances in quantum computing threaten the security of our systems and the communications we use. In order to face this threat, multiple solutions and protocols have been proposed in the Post-Quantum Cryptography project carried on by the National Institute of Standards and Technologies. The presented work focuses on defining a formal framework in Maude for the security analysis of different postquantum key encapsulation mechanisms under assumptions given under the DolevYao model. Through the use of our framework, we construct a symbolic model to represent the behaviour of each of the participants of the protocol in a network. We then conduct reachability analysis and find a man-in-the-middle attack in each of them and a design vulnerability in Bit Flipping Key Encapsulation. For both cases, we provide some insights on possible solutions. Then, we use the Maude Linear Temporal Logic model checker to extend the analysis of the symbolic system regarding security, liveness and fairness properties. Liveness and fairness properties hold while the security property does not due to the man-in-the-middle attack and the design vulnerability in Bit Flipping Key Encapsulation., Victor Garcia and Santiago Escobar were supported by the grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe and by the grant PCI2020-120708-2 funded by MICIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR. Kazuhiro Ogata was supported by JST SICORP Grant Number JPMJSC20C2, Japan. Sedat Akleylek was supported by TUBITAK under Grant No. 121R006. Ayoub Otmani was supported by FAVPQC project funded by CNRS and by the grant ANR-22-PETQ-0008 PQ-TLS funded by Agence Nationale de la Recherche (ANR) within France 2030 program. There was no additional external funding received for this study. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Safety Enforcement via Programmable Strategies in Maude

  • Alpuente Frasnedo, María|||0000-0002-9268-1178
  • Escobar Román, Santiago|||0000-0002-3550-4781
  • Sapiña-Sanchis, Julia|||0000-0003-2994-6986
  • Ballis, Demis
  • Galán-Pascual, Daniel
[EN] This work aims to provide a general mechanism for safety enforcement in rewriting logic computations. Our technique relies on an assertion-guided model transformation that leverages the newly defined Maude strategy language for ensuring rich safety policies in non-deterministic programs. The transformed system is guaranteed to comply with user -defined invariants that are expressed in a strategy-based, pattern-matching logic, thus preventing the concurrent system to reach any unsafe states. The performance and scal-ability of the technique is empirically evaluated and benchmarked on a set of realistic programs., This work was partially supported by TAILOR, a project funded by EU Horizon 2020 research and innovation programme under GA No 952215, grant PID2021-122830OB-C42 funded by MCIN/AEI/10.13039/501100011033 and by ?ERDF A way of making Europe?, and by Generalitat Valenciana under grant PROMETEO/2019/098.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

The impact of sociality regimes on heterogeneous cooperative-competitive multi-agent reinforcement learning: a study with the predator-prey game

  • Zhao, Yue
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] The performance in multi-agent reinforcement learning (MARL) scenarios has usually been analysed in homogeneous teams with a few choices for the sociality regime (selfish, egalitarian, or altruistic). In this paper we analyse both homogeneous and heterogeneous teams in a variation of sociality regimes in the predator-prey game, using a novel normalisation of the weights so that the sum of all rewards is independent of the sociality regime. We find that the selfish regime is advantageous for both predator and prey teams, and for both homogeneous and heterogeneous teams. In particular, rewards are about 100% higher for the predator team when switching from the egalitarian to selfish regime and more than 400% higher from the altruistic regime. For the prey, the increase is around 40% and 100% respectively. The results are similar for homogeneous and heterogeneous situations. The takeaway message is that any study of homogeneous and heterogeneous cooperative-competitive multi-agent reinforcement learning teams should also take into account the sociality regimes before making conclusions on the preference of any algorithm., This work was funded by the EU (FEDER) and Spanish grant RTI2018-094403-B-C32 funded by MCIN/AEI/10.13039/501100011033 and by 'ERDF A way of making Europe', Generalitat Valenciana under CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA), EU's Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR) and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and 'ERDF A way of making Europe' and China Scholarship Council (CSC) scholarship (No. 202006290201).
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Tesis doctoral (DoctoralThesis)

Contamination Problems in AI Evaluation

  • Mehrbakhsh, Behzad|||0000-0001-9017-989X
[ES] La evaluación de la Inteligencia Artificial se basa fundamentalmente en los benchmarks: conjuntos de datos y tareas estandarizadas que se utilizan para medir las capacidades de los modelos y orientar el progreso de la investigación. Sin embargo, la integridad de estas evaluaciones está cada vez más amenazada por la contaminación: distorsiones sutiles que comprometen la validez y la fiabilidad de las métricas de rendimiento. Esta tesis ofrece una investigación exhaustiva sobre la contaminación en la evaluación de sistemas de IA, analizando cómo surge, cómo afecta a los resultados reportados y cómo puede detectarse y mitigarse. Introducimos un marco teórico unificado que caracteriza la contaminación mediante ratios de densidad vecinal, cuantificando tanto la dirección como la magnitud del sesgo evaluativo. Dentro de este marco, identificamos y analizamos tres formas principales de contaminación, cada una con características e impactos distintos: La contaminación por benchmarks adversariales se produce cuando los conjuntos de evaluación se enriquecen deliberadamente con ejemplos difíciles. Aunque esta práctica es útil para pruebas de estrés y para revelar vulnerabilidades, desplaza sistemáticamente la distribución de dificultad respecto a las tareas del mundo real. Demostramos que las métricas de rendimiento agregadas confunden dos efectos distintos: el desplazamiento hacia ejemplos más difíciles y los cambios genuinos de capacidad. Mediante la descomposición del rendimiento a través de curvas características condicionadas por la dificultad, mostramos que el rendimiento sobre la distribución original puede recuperarse sin una medición directa, siempre que la probabilidad condicional respecto a la dificultad permanezca inalterada. La contaminación de datos surge cuando ejemplos del test o sus variantes cercanas aparecen en los datos de entrenamiento, inflando artificialmente el rendimiento mediante memorización en lugar de aprendizaje genuino. Replanteamos este fenómeno como un problema de asignación de recursos, introduciendo el concepto de presupuesto de contaminación para examinar los compromisos entre amplitud (número de ejemplos), profundidad (repeticiones por ejemplo) y dificultad (priorizar instancias más fáciles o más difíciles). A través de experimentos controlados con múltiples familias de modelos de lenguaje y benchmarks diversos, observamos que la magnitud de la contaminación no determina por sí sola la inflación del rendimiento: la estrategia de asignación es igualmente crítica. La contaminación generada por IA surge a medida que los modelos generativos pueblan Internet con contenido sintético que se vuelve indistinguible del material producido por humanos. Esta forma de contaminación es especialmente insidiosa porque se infiltra en los corpus de entrenamiento y en los benchmarks de evaluación mediante técnicas de web scraping. Nuestros experimentos controlados muestran que los errores generados por IA presentan características fundamentalmente distintas a los errores humanos: menor entropía, sesgos más sistemáticos y patrones de error diferenciados. Esta tesis avanza tanto la comprensión teórica como la detección práctica de la contaminación en la evaluación de sistemas de IA. Al hacer visible, cuantificable e interpretable la contaminación mediante herramientas como los ratios de densidad vecinal y las curvas características, proporcionamos una base para restaurar la confianza en la evaluación de la IA. A medida que avanzamos hacia sistemas cada vez más capaces, la integridad de nuestros métodos de evaluación nunca ha sido tan crítica. Este trabajo representa un paso hacia garantizar que el progreso reportado refleje avances genuinos y no artefactos de medición, permitiendo evaluaciones más fiables de las capacidades de la IA y decisiones de despliegue más seguras., [EN] Artificial Intelligence evaluation relies fundamentally on benchmarks; standardised datasets and tasks used to measure model capabilities and guide research progress. However, the integrity of these evaluations is increasingly threatened by contamination: subtle distortions that compromise the validity and reliability of performance assessments. This thesis provides a comprehensive investigation of contamination in AI evaluation, examining how it arises, how it affects reported performance, and how it can be detected and mitigated. We introduce a unified theoretical framework that characterises contamination through neighbourhood density ratios, quantifying both the direction and magnitude of evaluation bias. Within this framework, we identify and analyse three primary forms of contamination, each with distinct characteristics and impacts: Adversarial Benchmark Contamination occurs when evaluation sets are deliberately enriched with difficult examples. Although valuable for stress-testing and exposing weaknesses, this practice systematically shifts the difficulty distribution away from real-world tasks. We demonstrate that aggregate performance metrics confound two distinct effects: the shift toward harder examples and genuine capability changes. By decomposing performance through system characteristic curves conditioned on difficulty, we show that performance on the original distribution can be recovered without direct measurement, provided the conditional probability on difficulty remains unchanged. Data Contamination arises when test examples or their close variants appear in training data, artificially inflating performance through memorisation rather than genuine learning. We reframe this as a resource allocation problem, introducing the concept of a contamination budget to examine trade-offs between breadth (number of examples), depth (repetitions per example) and difficulty (prioritising easier or harder instances). Through controlled experiments across multiple language model families and diverse benchmarks, we find that contamination magnitude alone does not determine performance inflation, but the allocation strategy is equally critical. AI-Generated Contamination emerges as generative models increasingly populate the internet with synthetic content that becomes indistinguishable from human-authored material. This contamination is particularly insidious because it infiltrates both training corpora and evaluation benchmarks through web scraping. Our controlled experiments demonstrate that AI-generated errors exhibit fundamentally different characteristics than human errors: lower entropy, more systematic biases, and distinct error patterns. This thesis advances both the theoretical understanding and practical detection of evaluation contamination in AI systems. By making contamination visible, quantifiable, and interpretable through tools such as neighbourhood density ratios and characteristic curves, we provide a foundation for restoring trust in AI evaluation. As we advance toward increasingly capable AI systems, the integrity of our evaluation methods has never been more critical. This work represents a step toward ensuring that reported progress reflects genuine advances rather than measurement artefacts, enabling more reliable assessment of AI capabilities and safer deployment decisions., [CA] L'avaluació de la Intel·ligència Artificial es basa fonamentalment en els benchmarks: conjunts de dades i tasques estandarditzades que s'utilitzen per mesurar les capacitats dels models i orientar el progrés de la investigació. No obstant això, la integritat d'estes avaluacions està cada vegada més amenaçada per la contaminació: distorsions subtils que comprometen la validesa i la fiabilitat de les mètriques de rendiment. Esta tesi oferix una investigació exhaustiva sobre la contaminació en l'avaluació de sistemes d'IA, analitzant com apareix, com afecta els resultats reportats i com pot detectar-se i mitigar-se. Introduïm un marc teòric unificat que caracteritza la contaminació mitjançant ràtios de densitat veïnal, quantificant tant la direcció com la magnitud del biaix avaluatiu. Dins d'este marc, identifiquem i analitzem tres formes principals de contaminació, cadascuna amb característiques i impactes diferents: La contaminació per benchmarks adversarials es produïx quan els conjunts d'avaluació s'enriquixen deliberadament amb exemples difícils. Tot i que esta pràctica és útil per fer proves d'estrés i descobrir vulnerabilitats, desplaça sistemàticament la distribució de dificultat respecte a les tasques del món real. Demostrem que les mètriques de rendiment agregades confonen dos efectes diferents: el desplaçament cap a exemples més difícils i els canvis genuïns de capacitat. Mitjançant la descomposició del rendiment a través de corbes característiques condicionades per la dificultat, mostrem que el rendiment sobre la distribució original pot recuperar-se sense una mesura directa, sempre que la probabilitat condicional respecte a la dificultat es mantinga inalterada. La contaminació de dades sorgix quan exemples del test o variants pròximes apareixen en les dades d'entrenament, inflant artificialment el rendiment mitjançant memorització en lloc d'aprenentatge genuí. Replantegem este fenomen com un problema d'assignació de recursos, introduint el concepte de pressupost de contaminació per examinar els compromisos entre amplitud (nombre d'exemples), profunditat (repeticions per exemple) i dificultat (prioritzar instàncies més fàcils o més difícils). A través d'experiments controlats amb múltiples famílies de models de llenguatge i benchmarks diversos, observem que la magnitud de la contaminació no determina per si mateixa la inflació del rendiment: l'estratègia d'assignació és igualment crítica. La contaminació generada per IA apareix a mesura que els models generatius omplin Internet amb contingut sintètic que es torna indistinguible del material produït per humans. Esta forma de contaminació és especialment insidiosa perquè s'infiltra tant en els corpus d'entrenament com en els benchmarks d'avaluació mitjançant tècniques de web scraping. Els nostres experiments controlats mostren que els errors generats per IA presenten característiques fonamentalment diferents dels errors humans: menor entropia, biaixos més sistemàtics i patrons d'error diferenciats. Esta tesi avança tant la comprensió teòrica com la detecció pràctica de la contaminació en l'avaluació de sistemes d'IA. En fer visible, quantificable i interpretable la contaminació mitjançant eines com els ràtios de densitat veïnal i les corbes característiques, proporcionem una base per restaurar la confiança en l'avaluació de la IA. A mesura que avancem cap a sistemes cada vegada més capaços, la integritat dels nostres mètodes d'avaluació mai no ha sigut tan crítica. Este treball representa un pas cap a garantir que el progrés reportat reflectisca avanços genuïns i no artefactes de mesura, permetent avaluacions més fiables de les capacitats de la IA i decisions de desplegament més segures., This work was supported by several institutions and research projects. I gratefully acknowledge ValgrAI – Valencian Graduate School and Research Network for Artificial Intelligence and Generalitat Valenciana for providing predoctoral support. I also acknowledge funding from the EU’s Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR), which supported my research stay. This research was supported by the Spanish project PID2021-122830OB-C42 (SFERA), funded by MCIN/AEI/10.13039/501100011033 and ERDF “A way of making Europe”, and by the project CIPROM/2022/6 (FASSLOW), funded by Generalitat Valenciana – Programa Prometeo and ERDF, EU.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Comparative analysis of public and expert perceptions of electrified vehicles in the European Union

  • Mourtzouchou, Andromachi
  • Marin, Andres L.
  • Laveneziana, Lorenzo
  • Tansini, Alessandro
  • Suarez, Jaime
  • Garus, Ada
  • Grosso, Monica
  • Cordera, Rubén
  • Fontaras, Georgios
  • Ciuffo, Biagio
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
[EN] The transition to electric vehicles is widely regarded as an effective measure to combat climate change, and its success will utterly depend on the willingness of consumers to shift to this technology, making it essential to examine public perception when designing supportive policies. Nonetheless, public perception can be biased by multiple factors such as fake news, prejudices and misgivings, leading to a significant divergence from the views of experts in the transportation field. This paper examines the differences between the views of these two groups on electric vehicles, with a specific focus on the current perception in the European Union. The research adopts a two-fold approach, starting with an empirical driving study, followed by a series of focus group discussions with transportation experts and an analysis of Twitter (formally X) data. The data for this study was collected over five years, from June 2019 to June 2024. The findings indicate that both groups generally maintain a neutral stance, with neutrality appearing in 50% of their responses. However, experts tend to prefer battery electric vehicles, positively influenced by their driving experiences. In 40% of their comments, they report favourable sentiments toward these vehicles. Conversely, plug-in hybrid electric vehicles are favoured by the general public, with 35% of tweets indicating a positive attitude, as they are viewed as a more established technology. Differing perspectives on vehicle charging infrastructure are evident: specialists regard it as a hindrance to widespread adoption, with 30% expressing negative views, whereas the general population appears less concerned and more focused on the feelings and the experience of driving., This work was partly funded by CIPROM/2022/6 (FASSLOW) funded by Generalitat Valenciana, FISCALTICS (I+D+i PID2022-140110OA-I00) granted by MICIU/AEI/https://doi.org/10.13039/501100011033, and Spanis h grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/https://doi.org/10.13039/501100011033 and ERDF A way of making Europe .
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Analysis and discovery of procrastination patterns in a language learning MOOC

  • Vázquez Mendoza, Jorge A.
  • Ferri Ramírez, César|||0000-0002-8975-1120
  • C. Monserrat|||0000-0003-1790-8085
[EN] Online learning has been gaining a broad notoriety in society in the last years. The ease with which users from all over the world can learn is one of its main attractions. MOOCs are one of these technologies, which have enabled users to learn almost any subject of their interest. The use of MOOCs generates a massive amount of data that has been used by researchers with different goals: to predict dropout, predict grades and determine learning styles. Users of MOOCs face several challenges, some of which have been extensively studied. However, there are some of them, such as procrastination, whose study as a determining factor in the failure of such courses has not been addressed in sufficient depth. Through this paper, the influence of procrastination on the failure rates of users in a MOOC has been uncovered. Furthermore, by using process mining, this research has revealed the existence of a pattern of procrastination and the type of material used during study sessions by the users who fail the course. Speclifically, diverse forms of procrastination have been identified, resulting in differing effects on the educational outcomes attained by the user. These findings possess considerable implications, as they underscore the potential worth of detecting specific procrastination patterns to ascertain users who necessitate supplementary support during their educational pursuits within MOOCs., This work has been funded by ValGrai, CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana and PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10. 13039/501100011033 and "FEDER: a way to make Europe". Finally, we would like to thank the Department of Information and Communication Systems (ASIC) of the Polytechnic University of Valencia, with special mention to Ignacio Despujol, for providing us with access to the information and data used during the preparation of this article.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Contribución de congreso (ConferenceOutput)

Language Task Difficulty Prediction through LLM-Annotated Meta-Features

  • Moros-Daval, Yael
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Assessing the capabilities of large language models (LLMs) is increasingly challenging due to their generality and uneven task performance. Often, we do not know how much of the success or failure on a particular task is due to the `loading¿ of the language elements in the task, such as narrative understanding, or some other intrinsic (non-linguistic) components, such as domain-specific common sense or reasoning capabilities. Understanding what tasks are most loaded on language and determine the predictability of LLMs on these tasks is crucial for improving benchmarks, designing better LLMs, and ensuring their safe deployment. We present an innovative methodology that uses LLMs to annotate linguistic metafeatures , allowing us to predict task difficulty and understand linguistic loadings more accurately than traditional readability scores. Using GPT-4 for automated annotation, we show strong predictability for a variety of tasks and language models (e.g., MMLU with R2 from 0.68 to 0.83), but observe limited predictability for other tasks (e.g., LSAT with R2 of -0.07)., We acknowledge support from: Grant for Master Studies funded by ValgrAI Valencian Graduate School and Research Network for Artificial Intelligence and Generalitat Valenciana; FISCALTICS (I+D+i PID2022-140110OA-I00), granted by MICIU/AEI/10.13039/ 501100011033 and by ERDF, EU; CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana; the EC H2020-EU grant agreement No. 952215 (TAILOR); US DARPA HR00112120007 (RECoG-AI) and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ERDF.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Otros (Other)

Further Details on Examining Adversarial Evaluation: Role of Difficulty

  • Mehrbakhsh, Behzad
  • Hernández-Orallo, José
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
Adversarial benchmark construction, where harder instances challenge new generations of AI systems, is becoming the norm. While this approach may lead to better machine learning models ---on average and for the new \mbox{benchmark---,} it is unclear how these models behave on the original distribution. Two opposing effects are intertwined here. On the one hand, the adversarial benchmark has a higher proportion of difficult instances, with lower expected performance. On the other hand, models trained on the adversarial benchmark may improve on these difficult instances (but may also neglect some easy ones). To disentangle these two effects we can control for difficulty, showing that we can recover the performance on the original distribution, provided the harder instances were obtained from this distribution in the first place. We show this difficulty-aware rectification works in practice, through a series of experiments with several benchmark construction schemas and the use of a populational difficulty metric. As a take-away message, instead of distributional averages we recommend using difficulty-conditioned characteristic curves when evaluating models built with adversarial benchmarks., We thank the anonymous reviewers for their comments. This work was funded by valgrAI, the Norwegian Research Council grant 329745 Machine Teaching for Explainable AI, the Future of Life Institute, FLI, under grant RFP2-152, the EU (FEDER) and Spanish grant RTI2018-094403-B-C32 funded by MCIN/AEI/10.13039/501100011033 and by CIPROM/2022/6 funded by Generalitat Valenciana, EU’s Horizon 2020 research and innovation programme under grant agreement No. 952215 (TAILOR), US DARPA HR00112120007 (RECoG-AI) and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and "ERDF A way of making Europe" In compliance with the recommendations of the Science paper about reporting of evaluation results in AI [3], we include all the results at the instance level
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Conversational complexity for assessing risk in large language models

  • Burden, John
  • Cebrian, Manuel
  • Hernández-Orallo, José|||0000-0001-9746-7632
[EN] Large Language Models (LLMs) present a dual-use dilemma: they enable beneficial applications while harboring potential for harm, particularly through conversational interactions. Despite various safeguards, advanced LLMs remain vulnerable. A watersed case in early 2023 involved journalist Kevin Roose's extended dialogue with Bing, an LLM-powered search engine, which revealed harmful outputs after probing questions, highlighting vulnerabilities in the model's safeguards. This contrasts with simpler early jailbreaks, like the "Grandma Jailbreak," where users framed requests as innocent help for a grandmother, easily eliciting similar content. This raises the question: How much conversational effort is needed to elicit harmful information from LLMs? We propose two measures: Conversational Length (CL), which quantifies the conversation length used to obtain a specific response, and Conversational Complexity (CC), defined as the Kolmogorov complexity of the user's instruction sequence leading to the response. To address the incomputability of Kolmogorov complexity, we approximate CC using a reference LLM to estimate the compressibility of user instructions. Applying this approach to a large red-teaming dataset, we perform a quantitative analysis examining the statistical distribution of harmful and harmless conversational lengths and complexities. Our empirical findings suggest that this distributional analysis and the minimisation of CC serve as valuable tools for understanding AI safety, offering insights into the accessibility of harmful information. This work establishes a foundation for a new perspective on LLM safety, centered around the algorithmic complexity of pathways to harm., JB acknowledges support from Effective Ventures Foundation Long Term Future Fund Grant ID: a3rAJ000000017iYAA and US DARPA HR00112120007 (RECoG-AI) MC acknowledges support from multiple grants: PID2023-150271NB-C21 funded by the Ministerio de Ciencia, Innovación y Universidades, Agencia Estatal de Investigación; project PID2022-137243OB-I00 financed by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe ; and project TSI-100922-2023-0001 under the Convocatoria Cátedras ENIA 2022. Additional funding was provided by Google.org through the Silicon Valley Community Foundation via a grant to the Fundación General CSIC. Google.org had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. JHO thanks CIPROM/2022/6 (FASSLOW) and IDIFEDER/2021/05 (CLUSTERIA) funded by Generalitat Valenciana, the EC H2020-EU grant agreement No. 952215 (TAILOR), US DARPA HR00112120007 (RECoG-AI) and Spanish grant PID2021-122830OB-C42 (SFERA) funded by MCIN/AEI/10.13039/501100011033 and ERDF A way of making Europe .
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Human or Machine: A Novel Deep Learning Framework for Autonomous Driver Identification Based on Vehicle Trajectories

  • Laverde-Marin, Andres Felipe
  • Mattas, Konstantinos
  • Fontaras, Georgios
  • Kouvelas, Anastasios
  • Makridis, Michail A.
  • Martínez-Plumed, Fernando|||0000-0003-2902-6477
  • Ramírez Quintana, María José|||0000-0002-0559-3568
[EN] Monitoring traffic streams through vehicle trajectories offers valuable insights into traffic flow characteristics. In recent years, there has been a surge in the availability of vehicle trajectory datasets. At the same time, the number of autonomously-driven vehicles on the road is increasing, largely due to the adoption of systems like adaptive cruise control. However, distinguishing system-controlled vehicles from human-driven vehicles remains challenging, despite its potential to enable valuable applications and informed policy-making. The differences between the driving behavior of human-driven (HDs) and automated (ADs) vehicles in the longitudinal direction are highlighted in the literature and hold promise for novel methodologies that exploit them to identify the type of driver. Here, we propose a novel online-offline framework with three key contributions. First, a feature design component performs feature disentanglement to increase the performance of downstream deep learning models. Second, a bidirectional LSTM (bLSTM) architecture demonstrates excellent accuracy in differentiating between HD and AD vehicles. Third, a data drift detection component identifies changes in data distributions, enabling the framework to generalize effectively to unseen datasets with minimal new labeled observations., This work was supported in part by the Generalitat Valenciana under Grant CIPROM/2022/6 (FASSLOW) and Grant IDIFEDER/2021/05(CLUSTERIA), in part by European Commission (EC) H2020-European Union (EU) under Agreement 952215 (TAILOR), in part by Spanish funded by Ministerio de Ciencia e Innovacion (MCIN)/Agencia Estatal de Investigacion (AEI)/10.13039/501100011033 under Grant PID2021-122830OB-C42(SFERA), and in part by "European Regional Development Fund (ERDF) aWay of Making Europe."