Privacy-Aware Adaptive Model Splitting for Device-Edge Co Inference

The device-edge co-inference has emerged as a promising enabling technology for edge intelligence, which bal-ances the computational workload and communication overhead through model splitting. However, device-edge co-inference will expose the users' intermediate feature information to potential adversaries, and the adversaries may obtain the users' original data and location information through various possible means. In this paper, we present a privacy-aware adaptive model splitting approach to balance privacy preservation and communication-computation performance in device-edge co-inference. First, we use the reconstruction image similarity to evaluate the influence of reconstruction attacks on original data privacy under different splitting points. Second, we design an entropy-based location privacy evaluation metric to evaluate the location privacy of different splitting points, and propose a differential privacy-based transmit power perturbation mechanism to protect location privacy. Third, to balance the performance of privacy and communication-computation, we aim to solve the tradeoff problem of data privacy, location privacy, latency and energy consumption by optimizing the splitting point. Furthermore, we propose a deep reinforcement learning (DRL) based Adaptive Splitting Point Selection (DRL-ASPS) algorithm to solve the tradeoff problem. Simulations show that our proposed DRL-ASPS can effectively protect data and location privacy while maintaining the latency and energy consumption performance of co- inference.