Improving the transferability of adversarial examples against deep learning based computer networks

Convolutional Neural Networks (CNNs) are the most widely used algorithms in Deep Learning. The reason of their wide use are their effective architecture and high performance on classification detection tasks e.g., in the computer vision, image processing and especially in cybersecurity. Recent studies show that CNNs are easily attacked by the adversary, and they are open to vulnerabilities. Attacks on CNNs could be effective both on the targeted models and untargeted models (Nowroozi, Mekdad, Berenjestanaki, Conti & El Fergougui, 2022). We call this forceful property as a transferability. The transferability of the attacks also eliminates the need for the attacker to have perfect knowledge (PK) of the attacked network therefore attacker does not need to know the features, data and parameters (Nowroozi, Dehghantanha, Parizi & Choo, 2021). Recent studies prove that there is no transferability in computer networks. But in this study, a new strategy is carried out to improve the transferability of adversarial attacks between Source Network (SN) and Target Network (TN). In this strategy, we have modified an attack library to improve transferability by increasing the strength value of the attacks to make attacks stronger (Nowroozi, et.al., 2021). In this thesis, attacker allows a larger distortion to get an attacked sample which lies deeper in the target region, if the attack goes more inside the other class region the capability can improve which means the attack confidence become high. While investigating transferability issue between two networks, six attack types and two known datasets of computer networks are considered.