Dual-Timescale Adaptive Model Partitioning, Jamming Power Control, and Resource Allocation Against Multi-Eavesdroppers

Abstract

Model partitioning is a key technique for cooperative edge-device inference, balancing computing loads, reducing communication latency, and preserving privacy by transmitting intermediate parameters instead of raw data. However, eavesdroppers may intercept these intermediate parameters during wireless transmission, potentially recovering sensitive raw data. Current model partitioning studies primarily focus on latency and energy optimization, while neglecting security protection in multi-eavesdropper scenarios. To address this, we propose a dual-timescale adaptive model partitioning, jamming power control and resource allocation scheme against multi-eavesdroppers. Specifically, we jointly optimize model partition point selection for each device, jamming power control from the base station to each eavesdropper, computing resource allocation between end devices and the edge server for each tasks, and transmission power control for each device to minimize the overall latency-energy cost. To tackle the high complexity of the optimization problem and the heterogeneity in update frequencies among optimization variables, we propose a Twin Delayed Deep Deterministic Policy Gradient (TD3)-based dual-timescale reinforcement learning algorithm that jointly optimize the long-term model partitioning decision and other four short-term variables. By integrating numerical optimization techniques, the algorithm effectively reduces action space dimensionality and accelerates convergence. Extensive multi-scenario experiments demonstrate that our algorithm achieves the lowest total system cost and outperforms 4 existing advanced schemes.