Towards Deep Learning Models Resistant to Adversarial Attacks (PGD Training)

Formulated adversarial robustness as min-max robust optimization, demonstrating that Projected Gradient Descent (PGD) is the universal first-order adversary.

Serves as qualified reference for implementing task-adversarial-robustness in production systems.

Assumptions

• Underlying computational topology and mathematical bounds adhere to established convexity/smoothness guarantees

Limitations

• Hardware runtime speedups, privacy budgets, and convergence depend on hyperparameters and network communication limits