Black-Box Adversarial Attack for Deep Learning Classifiers in IoT Applications

Abstract

The increasing adoption of Internet of Things (IoT) has resulted in the availability of big data, which can reveal valuable insights if processed efficiently. Classification tasks are very important in such applications, and Artificial Intelligence is widely used to solve these problems. This paper demonstrates that Deep Learning classifiers used in IoT environments are vulnerable to black-box adversarial attacks. Such attacks can render these models ineffective by causing severe performance issues. This paper develops a black-box adversarial attack mechanism to generate adversarial examples for data obtained from smart meters installed in residential houses. An analysis is presented to demonstrate that the statistical properties of these adversarial examples are indistinguishable from those of the true examples, and the performance of the targeted models degrades sharply when exposed to the proposed attack. Further, the inherent properties of the attack mechanism imply that it is able to evade the entire class of gradient masking based defence methods. The effectiveness of the proposed black-box adversarial attack is demonstrated on the publicly available United Kingdom-Domestic Appliance-Level Electricity smart meter dataset.