Privacy-Preserving IoT Stream Analytics Using Homomorphic Encryption and Autoencoder-Based Anomaly Detection

Résumé fourni par la source

Over the last few years, the rapid evolution of Internet of Things (IoT) devices has resulted in the generation of a substantial number of sensitive data streams, demanding techniques that ensure both privacy preservation and real-time anomaly detection. However, the existing privacy-preserving anomaly detection approach, which integrates differential privacy with Cumulative Sum Algorithm (CUSUM)-based statistical detection, relies on simple statistical shifts and also faces sensitive information leakage due to exposed intermediate statistics. Hence, this research proposes a Privacy-Preserving Autoencoder (AE) framework integrating Homomorphic Encryption (HE) with data stream mining. Initially, continuous multivariate data streams are collected and preprocessed using moving-average smoothing, normalization, and sliding-window feature construction. Next, the preprocessed data are encrypted using Cheon-Kim-Kim-Song (CKKS) HE schemes with Single Instruction Multiple Data (SIMD) packing. Finally, anomaly detection is performed directly on encrypted data using the AE at the operator, where the reconstruction error identifies abnormal behavior in the data streams. Experimental results demonstrate the superior performance of the Autoencoder and Homomorphic Encryption (AE-HE)-based privacy-preserving anomaly detection framework with an Average Detection Delay (ADD), False Alarm Probability, and precision of six samples of 0.01 and 95 %, respectively.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.