Towards Secure Co-Inference: Leveraging DNN Partitioning and Collaborative Friendly Jamming

This paper proposes a new two-stage proximal policy optimization (PPO) algorithm, named TS-PPO, which jointly optimizes the End-to-End (E2E) delay and energy consumption of edge-enabled collaborative inference (co-inference) between an edge server (ES) and energy-harvesting users. The vulnerability of intermediate feature maps to the eavesdropping over wireless channels during offloading is considered and addressed through collaborative friendly jamming. Deep neural network (DNN) partitioning decisions are learned by the PPO in the first stage of TS-PPO, while closed-form solutions of computational resource allocation and a linear programming-based layered algorithm for friendly jamming are derived in the second stage. Experiments demonstrate that TS-PPO can improve the system reward by 53.3% compared to its benchmarks, while ensuring secure feature map transmissions.