EI-MTD: Moving Target Defense for Edge Intelligence against Adversarial Attacks

Edge intelligence has played an important role in constructing smart cities, but the vulnerability of edge nodes to adversarial attacks becomes urgent problem. A so-called example can fool a deep learning model on node for misclassification. Due transferability property examples, adversary easily black-box by local substitute model. general have limited resources, which cannot afford complicated defense mechanism like that cloud data center. To address challenge, we propose dynamic mechanism, namely EI-MTD. The first obtains robust member models small size through differential knowledge distillation from teacher Then, scheduling policy, builds Bayesian Stackelberg game, is applied choice target service. This prohibit selecting optimal attacks. We also conduct extensive experiments evaluate proposed and results show EI-MTD could protect effectively against settings.