RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publication
Artículo científico (JournalArticle)
Intrusion Detection Systems for the Internet of Drones Security Using Multi-Teacher Knowledge Distillation
[EN] The Internet of Drones (IoD) is increasingly adopted in mission-critical applications, but
it remains vulnerable to a wide range of cyberattacks due to its open communication channels and
resource-constrained nodes. Existing Intrusion Detection Systems (IDS) often neglect class imbalance and
computational limitations, resulting in poor generalization and impractical deployment. This study aims to
develop an efficient and deployable IDS framework for the IoD by addressing two key challenges: (i) the
severe class imbalance in real-world traffic datasets, and (ii) the high computational complexity of Deep
Learning (DL)-based models. We propose a two-phase methodology. In Phase 1, validated undersampling
strategies are applied to rebalance the CIC-IoT2023 dataset, after which baseline classifiers are trained
to assess detection performance. In Phase 2, the best-performing models serve as teacher networks in
a multi-teacher knowledge distillation (MTKD) framework¿implemented with average-weighted and
confidence-weighted variants¿to train lightweight Multi-Layer Perceptron (MLP) student models of
varying capacities. These students are optimized for deployment on resource-constrained IoD edge devices.
Effectiveness is assessed using both classification and computational efficiency metrics. Results show that
Instance Hardness Threshold undersampling consistently outperformed Random Under Sampler across all
baseline models, enhancing detection of minority attack classes. The best-performing models¿LSTM and
XGBoost¿were selected as teachers for MTKD. Applying average-weighted and confidence-weighted
MTKD to MLP1 × 8 increased accuracy by 0.03% and 0.09%, respectively, while confidence-weighted
also reduced inference time by 0.0006 s. These gains were achieved without increasing the computational
footprint, as the model retained its minimal size (0.0039 MB) and low parameter count (448). The proposed
two-phase framework provides a practical solution for building lightweight and accurate IDS models suitable
for real-time deployment in resource-constrained drone environments., This work was partially supported by research project MORELLINO (CIPROM/2023/29), funded by Direcció General de Ciència i
Investigació Generalitat Valenciana - SPAIN, and by the European Union s Horizon Europe project REMARKABLE under the
MSCA-SE grant ID: 101086387.