A Deep Dive into Deep Learning-Based Adversarial Attacks and Defenses in Computer Vision: From a Perspective of Cybersecurity

Adversarial attacks are deliberate data manipulations that may appear harmless to the viewer yet lead to incorrect categorization in a machine learning or deep learning system. These kinds of attacks frequently take shape of carefully constructed “noise”, that may cause misdiagnosis. Deep Neural Networks (DNNs) are being utilized more frequently in the physical world for applications that require high safety standards, including intelligent driverless vehicles, cloud systems, electronic health records, etc. The guiding rules of autonomous automobiles rely mainly on their capacity to gather environmental data through embedded sensors, then use DNN categorization to understand it and come to operational conclusions. For academics, the security and dependability of DNNs present numerous difficulties and worries. Threats of adversarial attacks on DNNs are one of the difficulties that researchers are now facing in their work. An adversarial attack assumes a minor alteration of an original image, with the modifications being nearly undetectable to human vision. In this survey, we describe the most recent studies on adversarial attack mechanisms on DNN-based frameworks and countermeasures to mitigate these attacks. Our main focus in this paper is images based adversarial attacks. We confer each research's benefits, drawbacks, methodology, and experimental results.