Advanced adversarial attack architecture for machine learning models

Adversarial attacks are defined as attacks aimed at deceiving machine learning systems. The small perturbations used in these attacks pose significant risks. This study presents the development of a PyTorch-based Generative Adversarial Network (GAN) architecture for generating adversarial attacks in the field of cybersecurity. The research includes a systematic comparison of four different optimization approaches: a baseline GAN model trained with Adam optimization, a genetic algorithm-based GAN approach distinguished by global search capabilities, a multi-objective evolutionary algorithm (MOEA) GAN strategy producing Pareto-optimal solutions, and an innovative hybrid approach combining the fast convergence advantages of Adam optimization with the global search capabilities of evolutionary algorithms. In experimental studies conducted against Random Forest classifier on the UNSW-NB15 dataset, the proposed hybrid approach demonstrated the highest attack effectiveness with 29.82% accuracy drop. The combination of genetic algorithm-based generator optimization with MOEA-based discriminator optimization achieved 633% higher attack success compared to traditional gradient-based methods. Attacks performed under L2 norm constraint successfully maintained the stealth-effectiveness balance. Results indicate that hybrid optimization strategies can establish a new paradigm in adversarial machine learning and open doors to significant developments in cybersecurity from both attack and defense perspectives.