Evaluation of GAN-Based Model for Adversarial Training.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models are susceptible to adversarial attacks, compromising their reliability.
• Generative Adversarial Networks (GANs) offer a potential solution for training robust classifiers.

Purpose of the Study:

• To present a novel GAN model for defending against L-infinity and L2 constrained gradient-based adversarial attacks.
• To address limitations of existing adversarial training methods, including gradient masking and complexity.

Main Methods:

• Developed a novel GAN with a dual generator architecture and new input formulations.
• Implemented unique L-infinity and L2 norm constraint vector outputs.
• Evaluated the impact of training epochs and GAN parameter settings.

Main Results:

• The proposed GAN model demonstrated defense against PGD L2 (60% accuracy) and PGD L-infinity (45% accuracy) perturbations.
• GANs successfully overcame gradient masking and generated effective data augmentation perturbations.
• Robustness transfer between constraints and a robustness-accuracy tradeoff were observed.

Conclusions:

• The novel GAN model provides effective defense against specific adversarial attacks.
• Optimal GAN adversarial training requires significant gradient information from the target classifier.
• Further research is needed to address overfitting and generalization issues.