Adaptative Perturbation Patterns: Realistic Adversarial Learning for Robust Intrusion Detection

Adversarial attacks pose a major threat to machine learning and the systems that rely on it. In cybersecurity domain, adversarial cyber-attack examples capable of evading detection are especially concerning. Nonetheless, an example generated for domain with tabular data must be realistic within domain. This work establishes fundamental constraint levels required achieve realism introduces Adaptative Perturbation Pattern Method (A2PM) fulfill these constraints in gray-box setting. A2PM relies pattern sequences independently adapted characteristics each class create valid coherent perturbations. The proposed method was evaluated case study two scenarios: Enterprise Internet Things (IoT) networks. Multilayer Perceptron (MLP) Random Forest (RF) classifiers were created regular training, using CIC-IDS2017 IoT-23 datasets. scenario, targeted untargeted performed against classifiers, compared original network traffic flows assess their realism. obtained results demonstrate provides scalable generation examples, which can advantageous both training attacks.