TinyML strategies for privacy-preserving and cyber threat multi-classification in edge-IoT networks

TinyML strategies for privacy-preserving and cyber threat multi-classification in edge-IoT networks

연구 분야: Safety

학회: Computing

Edge IoT and Machine Learning (ML) technologies have experienced rapid growth, enabling intelligent systems across diverse sectors such as healthcare, logistics, and Industry 4.0. However, the massive distributed data generated by IoT devices often contains sensitive information, raising critical security and privacy concerns. This underscores the need for real-time threat detection on resource-constrained devices. To address these challenges, this paper introduces a novel approach for threat detection based on Tiny Machine Learning (TinyML), a lightweight ML paradigm tailored for edge environments. The proposed approach applies a series of optimization strategies for both tree-based and neural-based models, including feature selection based on importance scoring, hyperparameter tuning, quantization, pruning, and sparsification. Experimental evaluations on four benchmark datasets, CICIoT2023, Edge-IIoTset, CICIDS2017, and RT-IoT2022, demonstrate that the optimized TinyML models remain competitive with conventional ML approaches, achieving accuracies exceeding 98%, while significantly reducing inference time and resource usage. To ensure data integrity and secure communication, cryptographic signatures are implemented using Libsodium and the EdDSA scheme. Additionally, approach robustness is assessed under adversarial conditions using Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), HopSkipJump, and Boundary techniques. These results highlight the potential of the proposed approach to enhance security and privacy in Edge-IoT networks, paving the way for scalable, efficient, and trustworthy edge intelligence solutions for real-time threat detection.