SE-DHPPM: A Secure and Explainable Deep Hybrid Privacy-Preserving Model for Edge Intelligence

Abstract

Edge Intelligence has emerged as a promising paradigm to facilitate real-time decision making in healthcare, Internet of Things (IoT) and intelligent surveillance applications. However, the issue of data privacy, secure communication, model interpretability, and low latency computation of computation in heterogeneous edge environments with limited resources, and non-independent and identically distributed (non-IID) data remains challenging. In this paper, a Secure and Explainable Deep Hybrid Privacy-Preserving Model for Edge Intelligence (SE-DHPPM) is proposed to solve these problems. The proposed framework combines federated learning for decentralized model training, differential privacy to protect sensitive information, blockchain for secure and tamper-resistant coordination, hybrid CNN-RNN architecture to learn the spatial-temporal feature, and explainable artificial intelligence (XAI) to improve the transparency of the prediction. The framework is tested with UNSW-NB15, CIC-IDS2017, MIMIC-III and WSN-DS benchmark datasets with realistic heterogeneous non-IID edge scenarios. Experimental results show that SE-DHPPM can achieve 96.8% accuracy, 96.2% precision, 95.9% recall, 96.0%AUC on the test set with communication overhead of only 430 MB and training time of 890 s. These findings show that SE-DHPPM is capable of providing a good compromise between privacy, security, interpretability, and computational efficiency, and is thus a feasible solution to next generation edge intelligence systems.