New Convex Training Method Boosts Neural Network Robustness

Researchers have developed an innovative training methodology for shallow neural networks aimed at significantly improving their resilience to adversarial attacks. The core of this approach involves applying Lipschitz regularization, which helps constrain the network's sensitivity to input perturbations. The key breakthrough is the transformation of a typically non-convex optimization problem into a convex one, enabling efficient solution to global optimality. This procedure can also be applied as a post-processing step to existing pre-trained networks, ensuring that the resulting model is at least as performant as its initial state while gaining enhanced robustness. Experimental results on real-world datasets demonstrate superior objective values and improved accuracy and robustness in adversarial settings compared to current methods.