Convex training of Lipschitz-regularized shallow neural networks
Researchers have developed a novel training method for shallow neural networks that enhances their resilience to adversarial attacks. This technique involves solving a non-convex Lipschitz-regularized training problem by first introducing a convex restriction that can be efficiently optimized to achieve a global optimum. The proposed approach can be applied as a post-processing step to existing pre-trained networks, ensuring the resulting network performs no worse than the initial one. Experiments on real-world datasets demonstrate that this convex training program yields networks with improved objective values, accuracy, and robustness against adversarial perturbations compared to current methods. AI
IMPACT This research offers a method to improve the security and reliability of neural networks against malicious inputs.