Impact of Adversarial Examples on Classification: A Comparative Study

Vision systems are nowadays based on smart cameras that are integrated into Industrial Control System (ICS) using Industrial Internet of Things (IIoT) principles. The information between smart cameras and/or vision systems and the reminder of ICS is exchanged using different communication protocols. This opens the possibilities for cyber-attacks by malicious adversaries and introduces cybersecurity related challenges. Cyber-attacks on communication between smart cameras and ICS can cause wrong decisions, such as misclassification of images, and system performance issues. Given the significant role of vision systems in industrial processes, timely detection of attacks on communication links between smart cameras and ICS is crucial to mitigate or avoid negative effects. However, adversaries often employ advanced and covert methods to generate stealthy attacks evading conventional detection techniques. One of the emerging threats facing vision systems is Adversarial Examples (AEs) specially crafted to deceive Deep Learning (DL) algorithms for classification of images. In this paper, we explore various types of AEs and their impact on classification models based on DL. Four different techniques are employed for designing AEs. The performance evaluation of the AEs is carried out using a real-world dataset.