Demystifying the Transferability of Adversarial Attacks in Computer Networks

Convolutional Neural Networks (CNNs) models are one of the most frequently used deep learning networks, and extensively in both academia industry. Recent studies demonstrated that adversarial attacks against such can maintain their effectiveness even when on other than targeted by attacker. This major property is known as transferability, makes CNNs ill-suited for security applications. In this paper, we provide first comprehensive study which assesses robustness CNN-based computer networks transferability. Furthermore, investigate whether transferability issue holds our experiments, consider five different attacks: Iterative Fast Gradient Method (I-FGSM), Jacobian-based Saliency Map (JSMA), Limited-memory Broyden Fletcher Goldfarb Shanno BFGS (L-BFGS), Projected Descent (PGD), DeepFool attack. Then, perform these three well-known datasets: Network-based Detection IoT (N-BaIoT) dataset, Domain Generating Algorithms (DGA) RIPE Atlas dataset. Our experimental results show clearly happens specific use cases I-FGSM, JSMA, LBFGS scenarios, attack success rate target network range from 63.00% to 100%. Finally, suggest two shielding strategies hinder considering Most Powerful Attacks (MPAs), mismatch LSTM architecture.