Pasar al contenido principal

Resultados relacionados

Resultados totales (Incluyendo duplicados): 24
Encontrada(s) 1 página(s)

UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Probabilistic Timing Estimates in Scenarios Under Testing Constraints

  • Vilardell Moreno, Sergi
  • Rossi, Francesco
  • Giordana, Gabriele
  • Serra, Isabel
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Cazorla Almeida, Francisco Javier
Measurement-based probabilistic (MBP) methods like Extreme Value Theory (EVT) and the Markov's Inequality have been exploited to derive probabilistic Worst-Case Execution Time (pWCET) estimates. Usually, the reliability and accuracy of pWCET techniques have been evaluated on medium to large sample sizes, N = [103, 105]. However, several works increasingly advocate for containing the cost of carrying out the test campaign by reducing the number of executions (i.e. the sample size) required by pWCET analysis. Specific scenarios, for example, impose inherent limitations on the collection of timing measurements due to cost and availability of appropriate testing facilities. In this work, we analyze the impact of small sample sizes on MBP. Our analysis shows that classical EVT models for tail estimation require a threshold that estimates where the tail of the distribution begins. In low sample scenarios, the uncertainty in determining this threshold can compromise the reliability of EVT estimates. We also assess the impact of small samples on RESTK, a time forecast method based on Markov's Inequality. Our results with synthetic data and representative kernels show that RESTK provides the best trade-off in terms of trustworthiness and tightness for small samples, partly due to not relying on the selection of any threshold, as opposed to EVT., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAF- EXPLAIN Project (www.safeexplain.eu), grant agreement num. 101069595. Authors also appreciate the support given to the Re- search Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Artículo científico (JournalArticle)

On Neural Networks Redundancy and Diversity for Their Use in Safety-Critical Systems

  • Brando, Axel
  • Serra, Isabel
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Cazorla Almeida, Francisco Javier
  • Perez Cerrolaza, Jon
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
An increasing number of critical functionalities integrated in embedded critical systems rely on deep learning (DL) technology. This article summarizes certain key aspects of DL’s intrinsic stochastic and training-data-dependent nature that are at odds with current domain-specific functional safety standards. We exemplify how redundancy and diversity of neural networks can help to reconcile DL technology and functional safety requirements., The research leading to these results has received funding from the European Research Council (ERC) grant agreement No. 772773 (SuPerCom), the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595, and the Spanish Ministry of Science and Innovation under grant PID2019-107255GBC21/AEI/10.13039/501100011033., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Semantic diverse redundancy for robust multi-modal safety-critical AI tasks

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
Dual Modular Redundancy (DMR) and Triple Modular Redundancy (TMR), often combined with diversity techniques, are widely used in safety-critical systems to achieve fault detection and/or tolerance. Traditional redundancy expects bit-identical outputs, flagging any mismatch as an error. However, emerging AI-based functionalities, such as camera and LiDAR-based object detection, are intrinsically stochastic and require only semantic correctness, allowing some variation in the outputs (e.g., slightly different confidence scores). In this work, we extend our previous semantic redundancy approach – originally developed for camera-based object detection – to LiDAR-based systems, which operate on 3D point cloud data. We propose software-only DMR and TMR schemes that introduce data-level diversity through domain-specific input transformations, preserving semantic meaning while increasing robustness. Our findings demonstrate that the method can be generalized to different AI tasks, helping improve safety and reliability in the different AI components of safety-critical systems., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work is also part of the project PLEC2023- 010240 funded by MICIU/AEI/10.13039/501100011033. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

SAFEXPLAIN: Safe and Explainable Critical Embedded Systems Based on AI

  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Perez, Jon
  • Englund, Cristofer
  • Zonooz, Bahram
  • Giordana, Gabriele
  • Cazorla Almeida, Francisco Javier
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Serra, Isabel
  • Brando, Axel
Deep Learning (DL) techniques are at the heart of most future advanced software functions in Critical Autonomous AI-based Systems (CAIS), where they also represent a major competitive factor. Hence, the economic success of CAIS industries (e.g., automotive, space, railway) depends on their ability to design, implement, qualify, and certify DL-based software products under bounded effort/cost. However, there is a fundamental gap between Functional Safety (FUSA) requirements on CAIS and the nature of DL solutions. This gap stems from the development process of DL libraries and affects high-level safety concepts such as (1) explainability and traceability, (2) suitability for varying safety requirements, (3) FUSA-compliant implementations, and (4) real-time constraints. As a matter of fact, the data-dependent and stochastic nature of DL algorithms clashes with current FUSA practice, which instead builds on deterministic, verifiable, and pass/fail test-based software. The SAFEXPLAIN project tackles these challenges and targets by providing a flexible approach to allow the certification - hence adoption - of DL-based solutions in CAIS building on: (1) DL solutions that provide end-to-end traceability, with specific approaches to explain whether predictions can be trusted and strategies to reach (and prove) correct operation, in accordance to certification standards; (2) alternative and increasingly sophisticated design safety patterns for DL with varying criticality and fault tolerance requirements; (3) DL library implementations that adhere to safety requirements; and (4) computing platform configurations, to regain determinism, and probabilistic timing analyses, to handle the remaining non-determinism., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. BSC authors have also been supported by the Spanish Ministry of Science and Innovation under grant PID2019- 107255GBC21/AEI/10.13039/501100011033., Peer Reviewed, "Article signat per 22 autors/es: Jaume Abella, Jon Perez, Cristofer Englund, Bahram Zonooz, Gabriele Giordana, Carlo Donzella, Francisco J. Cazorla, Enrico Mezzetti, Isabel Serra, Axel Brando, Irune Agirre, Fernando Eizaguirre, Thanh Hai Bui, Elahe Arani, Fahad Sarfraz, Ajay Balasubramaniam, Ahmed BadarIlaria Bloise, Lorenzo Feruglio, Ilaria Cinelli, Davide Brighenti, Davide Cunial"
UPCommons. Portal del coneixement obert de la UPC
Publicación Artículo científico (JournalArticle)

Assessing the use of NVIDIA multi-instance GPU in the automotive domain

  • Barrera Herrera, Javier Enrique|||0000-0003-1469-2514
  • Kosmidis, Leonidas
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Cazorla Almeida, Francisco Javier
The search for resource isolation and segregation is relentless to increase execution time determinism (ETD) as an essential feature for a successful certification process in embedded critical domains like automotive. In those domains, the advent of advanced software-controlled features requires unprecedented computing performance that calls for the use of acceleration hardware, with GPUs having a prominent position. The latest NVIDIA GPUs generations (Ampere, Hopper, Blackwell) feature MIG (Multi-Instance GPU), a mechanism that allows partitioning the GPU into fully isolated GPU instances, each with its own memory, cache, and computing resources. Despite the clear benefits of MIG on ETD, the latest NVIDIA automotive GPUs do not implement it. In this work, we first empirically analyze the benefits of MIG in a non-automotive GPU showing the main traits in its use to improve ETD. And second, we identify the potential reasons precluding the deployment of MIG for automotive GPUs: automotive market specific needs, and the difference between GPU and memory technologies used in high-performance GPUs, which implement MIG, and automotive GPUs that lack it., This work was supported in part by MICIU/AEI/10.13039/501100011033 under Project PLEC2023-010240; in part by the European Union’s HE Research and Innovation Programme under Grant 101069595; and in part by Agència de Gestió d’Ajuts Universitaris i de Recerca (AGAUR) under Grant 2021 SGR 00637. The work of Javier Barrera was supported by the “Generación D” initiative, Red.es, Ministerio para la Transformación Digital y de la Función Pública, for talent attraction, funded by the European Union NextGenerationEU funds, through PRTR under Grant C005/24-ED CV1. The work of Leonidas Kosmidis was supported by MICIU/AEI/10.13039/ 501100011033 under Grant IJC-2020- 045931-I., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

SAFEXPLAIN: a Complete Approach Towards Trustworthy AI-Based Safety-Critical Systems

  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Aguirre, Irune
  • Bui, Thanh
  • Geujen, Frank
  • Giordana, Gabriele
  • Cazorla Almeida, Francisco Javier
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Brando, Axel
Fault simulation and emulation are essential techniques for evaluating the dependability of integrated circuits, enabling early-stage vulnerability analysis and supporting the implementation of effective mitigation strategies. High-level hardware description languages such as Chisel facilitate the rapid development of complex fault scenarios with minimal modification to the design. However, existing Chisel-based fault injection (FI) frameworks are limited by their coarse-grained, instruction-level controllability, which restricts the precision of fault modeling. This work introduces Nail, a Chisel-based open-source FI framework that overcomes these limitations by introducing statebased faults. This approach allows fault scenarios based on specific system states instead of just instruction-level triggers, removing the need for precise timing of fault activation. For greater controllability, Nail allows users to arbitrarily modify internal trigger states via software at runtime. To support this, Nail automatically generates a software interface, offering straightforward access to the instrumented design. This enables fine-tuning of fault parameters during active fault-injection campaigns, a feature particularly beneficial for FPGA emulation, where synthesis is time-consuming. Utilizing these features, Nail narrows the gap between the high speed of emulation-based FI frameworks, the usability of software-based approaches, and the controllability achieved in simulation. We demonstrate Nail’s state-based fault injection and software framework by modeling a faulty general-purpose register in a RISC-V processor. Although this might appear straightforward, it requires statedependent fault injection and was previously impossible without fundamental changes to the design. The approach was validated in both simulation and FPGA emulation, where the addition of Nail introduced less than $1 \%$ resource overhead., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595., Peer Reviewed, "Article signat per 23 autors/es: Abella, Jaume and Agirre, Irune and Bui, Thanh and Geujen, Frank and Giordana, Gabriele and Donzella, Carlo and Cazorla, Francisco J. and Mezzetti, Enrico and Brando, Axel and Fernandez, Javier and Yarza, Irune and Plazaola, Joanes and Lowe, Robert and Ulan, Maria and Lavreysen, Rob and Tosi, Lucas and Bloise, Ilaria and Feruglio, Lorenzo and Cinelli, Ilaria and Lodico, Stefano and Guarienti, William and Nicosia, Giuseppe and Dallara, Valeria"
UPCommons. Portal del coneixement obert de la UPC
Publicación Artículo científico (JournalArticle)

Semantic diverse DMR and TMR for high-integrity AI-based function efficiency

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
Dual (DMR) and Triple Modular Redundancy (TMR), often with some form of diversity, are used in safety-critical systems to realize those functionalities at the highest integrity level providing fault detection and/or tolerance capabilities. Redundant executions are intended to provide bit-level identical results and, upon any mismatch, an error is assumed and recovery actions taken as needed. In this paper, we note that many emerging AI-based functionalities are intrinsically stochastic (e.g., camera-based object detection), and hence, their correctness must be judged semantically, with room for variations across correct outcomes (e.g., confidence must be above a given threshold, but how much it exceeds the threshold is irrelevant). Building on this observation, we propose strategies to create DMR and TMR implementations of AI-based functionalities that bring, not only fault tolerance against random hardware faults, but also against AI model inaccuracies. Those strategies, which can be realized with software-only means and ported to virtually any computing platform, build on input data modifications affecting the inference computations, but not the expected semantic output (e.g., introducing some controlled changes in the input data). Moreover, we provide our solution in the form of an open-source tool for image and video processing aimed at facilitating the reproducibility of our evaluation results, and enabling others to use it and conduct further research on input transformations., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work is also part of the project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033. Authors appreciate the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Safety-relevant AI-based system robustification with neural network ensembles

  • Aldomà Coll, Adrià|||0009-0005-1270-0295
  • Brando Guillaumes, Axel
  • Cazorla Almeida, Francisco Javier
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
Functional safety requirements of AI-based safety-critical applications challenge AI models, whose accuracy can be limited. In this paper, we show how using several cooperative deep learning (DL) models helps to raise global accuracy and reject making a prediction when confidence is below a pre-established threshold., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work has also been supported by the Project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033. Authors thank the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Leveraging image-based transformations to mitigate adversarial attacks in AI-based safety-critical systems

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. https://dx.doi.org/10.1109/IOLTS65288.2025.11116992, Dual (DMR) and Triple Modular Redundancy (TMR) are widely used techniques to provide fault detection and/or tolerance capabilities in safety-critical systems through – often diverse – redundancy. However, these systems remain vulnerable to adversarial attacks, which can mislead the AI models and lead to severe consequences. In this paper, we propose enhanced DMR and TMR implementations for image-based object detection leveraging image transformations during inference to mitigate the impact of adversarial attacks, hence addressing safety and security concerns simultaneously. Our approach achieves up to 12.9% and 12.2% higher accuracy in adversarial scenarios compared to state-of-the-art solutions in DMR and TMR configurations, respectively., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595. This work is also part of the project PLEC2023-010240 funded by ICIU/AEI/10.13039/501100011033. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Retrospective Uncertainties for Deep Models using Vine Copulas

  • Tagasovska, Natasa
  • Ozdemir, Firat
  • Brando Guillaumes, Axel
Despite the major progress of deep models as learning machines, uncertainty estimation remains a major challenge. Existing solutions rely on modified loss functions or architectural changes. We propose to compensate for the lack of built-in uncertainty estimates by supplementing any network, retrospectively, with a subsequent vine copula model, in an overall compound we call Vine-Copula Neural Network (VCNN). Through synthetic and real-data experiments, we show that VCNNs could be task (regression/classification) and architecture (recurrent, fully connected) agnostic while providing reliable and better-calibrated uncertainty estimates, comparable to state-of-the-art built-in uncertainty solutions., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 772773). Additionally, this work has been partially supported by Grant PID2019-107255GB-C21 funded by MCIN/AEI/ 10.13039/501100011033., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Detecting low-density mixtures in high-quantile tails for pWCET estimation

  • Manau Corderroure, Blau
  • Vilardell Moreno, Sergi
  • Serra Mochales, Isabel
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
  • Cazorla Almeida, Francisco Javier
The variability arising from sophisticated hardware and software solutions in cutting-edge embedded products causes software to exhibit complex execution time distributions. Mixture distributions can happen, with different density (weight), as a result of inherent different features in the execution platform and multiple operational scenarios. In the context of probabilistic WCET (pWCET) analysis based on Extreme Value Theory (EVT), where identical distribution is a pre-requisite, mixtures are typically intercepted by applying stationarity tests on the full sample. Those tests, however, are instructed to detect only mixtures with sufficiently high probability (weight) and disregard low-density mixtures (which are unlikely to be preserved in the high-quantile tail of the sample) as they would prevent any form of stationarity. Nonetheless, low-density mixture distributions can persist and even exacerbate in the tail, and, when not considered, they can impair pWCET estimation in EVT-based approaches, leading to overly pessimistic or optimistic bounds. In this work, we propose TailID, an iterative point-wise approach that builds on the asymptotic convergence of the Maximum Likelihood Estimator (MLE) of the Extreme Value Index (EVI) parameter ξ to detect low-density mixture distributions on high-quantile tails and use this information to steer EVT tail selection. The benefits of the proposed method are assessed on synthetic mixture distributions and real data collected on an industrially representative embedded platform., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033 (CAPSUL-IA). Authors also appreciate the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya. This work is partially supported by the Spanish MICIU with project PID2022-137414OB-I00., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publicación Contribución de congreso (ConferenceOutput)

Standardizing the Probabilistic Sources of Uncertainty for the sake of Safety Deep Learning

  • Brando, Axel
  • Serra, Isabel
  • Mezzetti, Enrico|||0000-0002-1886-2931
  • Cazorla Almeida, Francisco Javier
  • Abella Ferrer, Jaume|||0000-0001-7951-4028
Nowadays, critical functionalities are increasingly tackled by autonomous decision-making systems, which depend on Artificial Intelligence (e.g. Deep Learning) models. Still, most of these models are designed to maximize the generic performance rather than preventing potential irreversible errors. While robustness and reliability techniques have been developed, in the recent years, to fill this gap, the sources of uncertainty in those decision models are still ambiguous. With a view to standardizing the uncertainty sources, in this paper we present a formal methodology to disentangle those sources from a probabilistic viewpoint for any (regression and classification) supervised learning model. Once we associate a formula to each uncertainty type, we expose the terminology disagreement in the literature and we propose one that is aligned with other previous works. Finally, based on the proposed formulation, we present an integrated visualization method to represent all the uncertainty sources in a single figure to, ultimately, assisting the design of uncertainty-tailored actions, The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 772773). Additionally, this work has been partially supported by Grant PID2019-107255GB-C21 funded by MCIN/AEI/ 10.13039/501100011033., Peer Reviewed
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

SAFEXPLAIN: a Complete Approach Towards Trustworthy AI-Based Safety-Critical Systems

  • Abella Ferrer, Jaume
  • Aguirre, Irune
  • Bui, Thanh
  • Geujen, Frank
  • Giordana, Gabriele
  • Cazorla Almeida, Francisco Javier
  • Mezzetti, Enrico
  • Brando, Axel
Fault simulation and emulation are essential techniques for evaluating the dependability of integrated circuits, enabling early-stage vulnerability analysis and supporting the implementation of effective mitigation strategies. High-level hardware description languages such as Chisel facilitate the rapid development of complex fault scenarios with minimal modification to the design. However, existing Chisel-based fault injection (FI) frameworks are limited by their coarse-grained, instruction-level controllability, which restricts the precision of fault modeling. This work introduces Nail, a Chisel-based open-source FI framework that overcomes these limitations by introducing statebased faults. This approach allows fault scenarios based on specific system states instead of just instruction-level triggers, removing the need for precise timing of fault activation. For greater controllability, Nail allows users to arbitrarily modify internal trigger states via software at runtime. To support this, Nail automatically generates a software interface, offering straightforward access to the instrumented design. This enables fine-tuning of fault parameters during active fault-injection campaigns, a feature particularly beneficial for FPGA emulation, where synthesis is time-consuming. Utilizing these features, Nail narrows the gap between the high speed of emulation-based FI frameworks, the usability of software-based approaches, and the controllability achieved in simulation. We demonstrate Nail’s state-based fault injection and software framework by modeling a faulty general-purpose register in a RISC-V processor. Although this might appear straightforward, it requires statedependent fault injection and was previously impossible without fundamental changes to the design. The approach was validated in both simulation and FPGA emulation, where the addition of Nail introduced less than $1 \%$ resource overhead., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595., Peer Reviewed, "Article signat per 23 autors/es: Abella, Jaume and Agirre, Irune and Bui, Thanh and Geujen, Frank and Giordana, Gabriele and Donzella, Carlo and Cazorla, Francisco J. and Mezzetti, Enrico and Brando, Axel and Fernandez, Javier and Yarza, Irune and Plazaola, Joanes and Lowe, Robert and Ulan, Maria and Lavreysen, Rob and Tosi, Lucas and Bloise, Ilaria and Feruglio, Lorenzo and Cinelli, Ilaria and Lodico, Stefano and Guarienti, William and Nicosia, Giuseppe and Dallara, Valeria", Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Semantic diverse redundancy for robust multi-modal safety-critical AI tasks

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume
Dual Modular Redundancy (DMR) and Triple Modular Redundancy (TMR), often combined with diversity techniques, are widely used in safety-critical systems to achieve fault detection and/or tolerance. Traditional redundancy expects bit-identical outputs, flagging any mismatch as an error. However, emerging AI-based functionalities, such as camera and LiDAR-based object detection, are intrinsically stochastic and require only semantic correctness, allowing some variation in the outputs (e.g., slightly different confidence scores). In this work, we extend our previous semantic redundancy approach – originally developed for camera-based object detection – to LiDAR-based systems, which operate on 3D point cloud data. We propose software-only DMR and TMR schemes that introduce data-level diversity through domain-specific input transformations, preserving semantic meaning while increasing robustness. Our findings demonstrate that the method can be generalized to different AI tasks, helping improve safety and reliability in the different AI components of safety-critical systems., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work is also part of the project PLEC2023- 010240 funded by MICIU/AEI/10.13039/501100011033. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

SAFEXPLAIN: Safe and Explainable Critical Embedded Systems Based on AI

  • Abella Ferrer, Jaume
  • Perez, Jon
  • Englund, Cristofer
  • Zonooz, Bahram
  • Giordana, Gabriele
  • Cazorla Almeida, Francisco Javier
  • Mezzetti, Enrico
  • Serra, Isabel
  • Brando, Axel
Deep Learning (DL) techniques are at the heart of most future advanced software functions in Critical Autonomous AI-based Systems (CAIS), where they also represent a major competitive factor. Hence, the economic success of CAIS industries (e.g., automotive, space, railway) depends on their ability to design, implement, qualify, and certify DL-based software products under bounded effort/cost. However, there is a fundamental gap between Functional Safety (FUSA) requirements on CAIS and the nature of DL solutions. This gap stems from the development process of DL libraries and affects high-level safety concepts such as (1) explainability and traceability, (2) suitability for varying safety requirements, (3) FUSA-compliant implementations, and (4) real-time constraints. As a matter of fact, the data-dependent and stochastic nature of DL algorithms clashes with current FUSA practice, which instead builds on deterministic, verifiable, and pass/fail test-based software. The SAFEXPLAIN project tackles these challenges and targets by providing a flexible approach to allow the certification - hence adoption - of DL-based solutions in CAIS building on: (1) DL solutions that provide end-to-end traceability, with specific approaches to explain whether predictions can be trusted and strategies to reach (and prove) correct operation, in accordance to certification standards; (2) alternative and increasingly sophisticated design safety patterns for DL with varying criticality and fault tolerance requirements; (3) DL library implementations that adhere to safety requirements; and (4) computing platform configurations, to regain determinism, and probabilistic timing analyses, to handle the remaining non-determinism., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. BSC authors have also been supported by the Spanish Ministry of Science and Innovation under grant PID2019- 107255GBC21/AEI/10.13039/501100011033., Peer Reviewed, "Article signat per 22 autors/es: Jaume Abella, Jon Perez, Cristofer Englund, Bahram Zonooz, Gabriele Giordana, Carlo Donzella, Francisco J. Cazorla, Enrico Mezzetti, Isabel Serra, Axel Brando, Irune Agirre, Fernando Eizaguirre, Thanh Hai Bui, Elahe Arani, Fahad Sarfraz, Ajay Balasubramaniam, Ahmed BadarIlaria Bloise, Lorenzo Feruglio, Ilaria Cinelli, Davide Brighenti, Davide Cunial", Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Retrospective Uncertainties for Deep Models using Vine Copulas

  • Tagasovska, Natasa
  • Ozdemir, Firat
  • Brando Guillaumes, Axel
Despite the major progress of deep models as learning machines, uncertainty estimation remains a major challenge. Existing solutions rely on modified loss functions or architectural changes. We propose to compensate for the lack of built-in uncertainty estimates by supplementing any network, retrospectively, with a subsequent vine copula model, in an overall compound we call Vine-Copula Neural Network (VCNN). Through synthetic and real-data experiments, we show that VCNNs could be task (regression/classification) and architecture (recurrent, fully connected) agnostic while providing reliable and better-calibrated uncertainty estimates, comparable to state-of-the-art built-in uncertainty solutions., The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 772773). Additionally, this work has been partially supported by Grant PID2019-107255GB-C21 funded by MCIN/AEI/ 10.13039/501100011033., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

On Neural Networks Redundancy and Diversity for Their Use in Safety-Critical Systems

  • Brando, Axel
  • Serra, Isabel
  • Mezzetti, Enrico
  • Cazorla Almeida, Francisco Javier
  • Perez Cerrolaza, Jon
  • Abella Ferrer, Jaume
An increasing number of critical functionalities integrated in embedded critical systems rely on deep learning (DL) technology. This article summarizes certain key aspects of DL’s intrinsic stochastic and training-data-dependent nature that are at odds with current domain-specific functional safety standards. We exemplify how redundancy and diversity of neural networks can help to reconcile DL technology and functional safety requirements., The research leading to these results has received funding from the European Research Council (ERC) grant agreement No. 772773 (SuPerCom), the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595, and the Spanish Ministry of Science and Innovation under grant PID2019-107255GBC21/AEI/10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Standardizing the Probabilistic Sources of Uncertainty for the sake of Safety Deep Learning

  • Brando, Axel
  • Serra, Isabel
  • Mezzetti, Enrico
  • Cazorla Almeida, Francisco Javier
  • Abella Ferrer, Jaume
Nowadays, critical functionalities are increasingly tackled by autonomous decision-making systems, which depend on Artificial Intelligence (e.g. Deep Learning) models. Still, most of these models are designed to maximize the generic performance rather than preventing potential irreversible errors. While robustness and reliability techniques have been developed, in the recent years, to fill this gap, the sources of uncertainty in those decision models are still ambiguous. With a view to standardizing the uncertainty sources, in this paper we present a formal methodology to disentangle those sources from a probabilistic viewpoint for any (regression and classification) supervised learning model. Once we associate a formula to each uncertainty type, we expose the terminology disagreement in the literature and we propose one that is aligned with other previous works. Finally, based on the proposed formulation, we present an integrated visualization method to represent all the uncertainty sources in a single figure to, ultimately, assisting the design of uncertainty-tailored actions, The research leading to these results has received funding from the Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 772773). Additionally, this work has been partially supported by Grant PID2019-107255GB-C21 funded by MCIN/AEI/ 10.13039/501100011033., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Semantic diverse DMR and TMR for high-integrity AI-based function efficiency

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume
Dual (DMR) and Triple Modular Redundancy (TMR), often with some form of diversity, are used in safety-critical systems to realize those functionalities at the highest integrity level providing fault detection and/or tolerance capabilities. Redundant executions are intended to provide bit-level identical results and, upon any mismatch, an error is assumed and recovery actions taken as needed. In this paper, we note that many emerging AI-based functionalities are intrinsically stochastic (e.g., camera-based object detection), and hence, their correctness must be judged semantically, with room for variations across correct outcomes (e.g., confidence must be above a given threshold, but how much it exceeds the threshold is irrelevant). Building on this observation, we propose strategies to create DMR and TMR implementations of AI-based functionalities that bring, not only fault tolerance against random hardware faults, but also against AI model inaccuracies. Those strategies, which can be realized with software-only means and ported to virtually any computing platform, build on input data modifications affecting the inference computations, but not the expected semantic output (e.g., introducing some controlled changes in the input data). Moreover, we provide our solution in the form of an open-source tool for image and video processing aimed at facilitating the reproducibility of our evaluation results, and enabling others to use it and conduct further research on input transformations., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work is also part of the project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033. Authors appreciate the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Assessing the use of NVIDIA multi-instance GPU in the automotive domain

  • Barrera Herrera, Javier Enrique
  • Kosmidis, Leonidas
  • Abella Ferrer, Jaume
  • Cazorla Almeida, Francisco Javier
The search for resource isolation and segregation is relentless to increase execution time determinism (ETD) as an essential feature for a successful certification process in embedded critical domains like automotive. In those domains, the advent of advanced software-controlled features requires unprecedented computing performance that calls for the use of acceleration hardware, with GPUs having a prominent position. The latest NVIDIA GPUs generations (Ampere, Hopper, Blackwell) feature MIG (Multi-Instance GPU), a mechanism that allows partitioning the GPU into fully isolated GPU instances, each with its own memory, cache, and computing resources. Despite the clear benefits of MIG on ETD, the latest NVIDIA automotive GPUs do not implement it. In this work, we first empirically analyze the benefits of MIG in a non-automotive GPU showing the main traits in its use to improve ETD. And second, we identify the potential reasons precluding the deployment of MIG for automotive GPUs: automotive market specific needs, and the difference between GPU and memory technologies used in high-performance GPUs, which implement MIG, and automotive GPUs that lack it., This work was supported in part by MICIU/AEI/10.13039/501100011033 under Project PLEC2023-010240; in part by the European Union’s HE Research and Innovation Programme under Grant 101069595; and in part by Agència de Gestió d’Ajuts Universitaris i de Recerca (AGAUR) under Grant 2021 SGR 00637. The work of Javier Barrera was supported by the “Generación D” initiative, Red.es, Ministerio para la Transformación Digital y de la Función Pública, for talent attraction, funded by the European Union NextGenerationEU funds, through PRTR under Grant C005/24-ED CV1. The work of Leonidas Kosmidis was supported by MICIU/AEI/10.13039/ 501100011033 under Grant IJC-2020- 045931-I., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Leveraging image-based transformations to mitigate adversarial attacks in AI-based safety-critical systems

  • Caro Roca, Martí
  • Brando Guillaumes, Axel
  • Abella Ferrer, Jaume
© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. https://dx.doi.org/10.1109/IOLTS65288.2025.11116992, Dual (DMR) and Triple Modular Redundancy (TMR) are widely used techniques to provide fault detection and/or tolerance capabilities in safety-critical systems through – often diverse – redundancy. However, these systems remain vulnerable to adversarial attacks, which can mislead the AI models and lead to severe consequences. In this paper, we propose enhanced DMR and TMR implementations for image-based object detection leveraging image transformations during inference to mitigate the impact of adversarial attacks, hence addressing safety and security concerns simultaneously. Our approach achieves up to 12.9% and 12.2% higher accuracy in adversarial scenarios compared to state-of-the-art solutions in DMR and TMR configurations, respectively., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num.101069595. This work is also part of the project PLEC2023-010240 funded by ICIU/AEI/10.13039/501100011033. M. Caro has received funding under grant PRE2022-102440 funded by MICIU/AEI/10.13039/501100011033 and, by ESF+., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Detecting low-density mixtures in high-quantile tails for pWCET estimation

  • Manau Corderroure, Blau
  • Vilardell Moreno, Sergi
  • Serra Mochales, Isabel
  • Mezzetti, Enrico
  • Abella Ferrer, Jaume
  • Cazorla Almeida, Francisco Javier
The variability arising from sophisticated hardware and software solutions in cutting-edge embedded products causes software to exhibit complex execution time distributions. Mixture distributions can happen, with different density (weight), as a result of inherent different features in the execution platform and multiple operational scenarios. In the context of probabilistic WCET (pWCET) analysis based on Extreme Value Theory (EVT), where identical distribution is a pre-requisite, mixtures are typically intercepted by applying stationarity tests on the full sample. Those tests, however, are instructed to detect only mixtures with sufficiently high probability (weight) and disregard low-density mixtures (which are unlikely to be preserved in the high-quantile tail of the sample) as they would prevent any form of stationarity. Nonetheless, low-density mixture distributions can persist and even exacerbate in the tail, and, when not considered, they can impair pWCET estimation in EVT-based approaches, leading to overly pessimistic or optimistic bounds. In this work, we propose TailID, an iterative point-wise approach that builds on the asymptotic convergence of the Maximum Likelihood Estimator (MLE) of the Extreme Value Index (EVI) parameter ξ to detect low-density mixture distributions on high-quantile tails and use this information to steer EVT tail selection. The benefits of the proposed method are assessed on synthetic mixture distributions and real data collected on an industrially representative embedded platform., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595 and project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033 (CAPSUL-IA). Authors also appreciate the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya. This work is partially supported by the Spanish MICIU with project PID2022-137414OB-I00., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Safety-relevant AI-based system robustification with neural network ensembles

  • Aldomà Coll, Adrià
  • Brando Guillaumes, Axel
  • Cazorla Almeida, Francisco Javier
  • Abella Ferrer, Jaume
Functional safety requirements of AI-based safety-critical applications challenge AI models, whose accuracy can be limited. In this paper, we show how using several cooperative deep learning (DL) models helps to raise global accuracy and reject making a prediction when confidence is below a pre-established threshold., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAFEXPLAIN Project (www.safexplain.eu), grant agreement num. 101069595. This work has also been supported by the Project PLEC2023-010240 funded by MICIU/AEI/10.13039/501100011033. Authors thank the support given to the Research Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publicación Contribución de congreso (ConferenceOutput)

Probabilistic Timing Estimates in Scenarios Under Testing Constraints

  • Vilardell Moreno, Sergi
  • Rossi, Francesco
  • Giordana, Gabriele
  • Serra, Isabel
  • Mezzetti, Enrico
  • Abella Ferrer, Jaume
  • Cazorla Almeida, Francisco Javier
Measurement-based probabilistic (MBP) methods like Extreme Value Theory (EVT) and the Markov's Inequality have been exploited to derive probabilistic Worst-Case Execution Time (pWCET) estimates. Usually, the reliability and accuracy of pWCET techniques have been evaluated on medium to large sample sizes, N = [103, 105]. However, several works increasingly advocate for containing the cost of carrying out the test campaign by reducing the number of executions (i.e. the sample size) required by pWCET analysis. Specific scenarios, for example, impose inherent limitations on the collection of timing measurements due to cost and availability of appropriate testing facilities. In this work, we analyze the impact of small sample sizes on MBP. Our analysis shows that classical EVT models for tail estimation require a threshold that estimates where the tail of the distribution begins. In low sample scenarios, the uncertainty in determining this threshold can compromise the reliability of EVT estimates. We also assess the impact of small samples on RESTK, a time forecast method based on Markov's Inequality. Our results with synthetic data and representative kernels show that RESTK provides the best trade-off in terms of trustworthiness and tightness for small samples, partly due to not relying on the selection of any threshold, as opposed to EVT., The research leading to these results has received funding from the European Union’s Horizon Europe Programme under the SAF- EXPLAIN Project (www.safeexplain.eu), grant agreement num. 101069595. Authors also appreciate the support given to the Re- search Group SSAS (Code: 2021 SGR 00637) by the Research and University Department of the Generalitat de Catalunya., Peer Reviewed, Postprint (author's final draft)