Exploiting the Divergence Between Output of ML Models to Detect Adversarial Attacks in Streaming IoT Applications

The majority of streaming Internet of Things (IoT) applications use machine learning models to identify and classify streaming inputs before forwarding them for further processing. These streaming IoT systems, however, are vulnerable to poisoning and adversarial attacks. An adversary deliberately modifies the input by adding a small perturbation during the communication to fool the class label into producing an arbitrary or specific output. The increasing number of well-developed, imperceptible attacks necessitates more sophisticated countermeasures. To this end, this paper underlines this problem and proposes a new scheme based on committee-based machine learning models: some have experience with only benign inputs, and others with benign and adversarial inputs. Then, the probabilities of the outputs of these pairs' models are utilized. The KL-divergence after that is applied to identify, detect, and mitigate such streaming attacks. Specifically, we use the uncertainty measures between the output of mitigation and non-mitigation ML models as a proxy to identify adversely attacked inputs. We use traffic sign classification in autonomous vehicle technology as a streaming IoT application. Our experiments demonstrate that the proposed approach can detect and mitigate adversarial attacks with high confidence for the white-box attack.