Reinforcement Learning-Based Efficient Multi-Exit Neural Networks Against Side-Channel Attacks

Distributed multi-exit neural networks (MeNNs) enable mobile devices to handle complex tasks such as image classification, but their performance is highly dependent on transmission quality and is therefore vulnerable to side-channel attacks. In this paper, we design a side-channel attack model and propose an efficient inference framework based on the distributed MeNN to resist the designed attack. First, we design an intelligent side-channel attack model, in which the attacker can eavesdrop on the communication channel and use deep reinforcement learning (RL) to predict the early exit decision of each sample. Next, we develop a defense method that employs a hierarchical and multi-agent RL to determine whether to infer locally or offload to a chosen early exit on the server, and to adjust the transmit power accordingly. We further propose a critic-guided safety mechanism that steers local agents away from risky policies that would cause inference failures or severe data leakage. We prove that our framework enforces a strict instantaneous security constraint and asymptotically achieves the optimum by deriving a regret bound. Extensive experiments on several datasets (including CIFAR‑10, CIFAR‑100, STL‑10, EMNIST, FMNIST, and Stanford Cars) show that our method reduces inference latency, improves classification accuracy, and significantly enhances robustness against side-channel attacks, as compared with two benchmarks SCAN and PCE.