Adversarial Robustness of NTK Neural Networks

Collection of recent ML research with direct relevance to security and safety in ML systems. Key items: (1) Adversarial Robustness of NTK Neural Networks — establishes minimax rates for adversarial regression, shows NTK-trained networks can achieve optimal robustness but proves the minimum-norm interpolant (overfitting regime) is vulnerable to adversarial perturbations. (2) Occam's Razor / ELBO — demonstrates that ELBO-based hyperparameter learning and rank-restricted approximate posteriors can produce overfitting or underfitting, causing model-selection failures. (3) Robust representation &Ge