Training and Verifying Robust Kolmogorov-Arnold Networks

Kolmogorov-Arnold Networks (KANs) offer strong theoretical representational power but, like MLPs and CNNs, remain vulnerable to adversarial attacks. Benchmarks on Fashion MNIST and CIFAR10 confirm this susceptibility. We introduce GloroKAN, leveraging KANs' B-spline structure to approximate local Lipschitz constants directly in the forward pass, boosting robustness without gradient-based adversarial training and approaching adversarially trained performance. Additionally, we propose a verification method based on propagating intervals. While these findings highlight KANs' potential for robust, interpretable models, further research is needed to realize their full promise.