Stability and Robustness via Regularization: Bandit Inference via Regularized Stochastic Mirror Descent

Collection of ML/optimization papers with several security-relevant implications. Key points: (1) “Stability and Robustness via Regularization” shows regularized stochastic mirror-descent (regularized-EXP3) yields statistical stability and provable robustness to o(T^{1/2}) adversarial corruptions, while common algorithms (e.g., UCB) can fail under small corruption—implications for deployed adaptive sampling/bandit systems and poisoning/online-corruption attacks. (2) Policy-gradient diffusion results identify learning-rate regimes where regret becomes linear, indicating that misconfiguration or