Robust Learning for Data Poisoning Attacks

We investigate the robustness of stochastic approximation approaches against data poisoning attacks. We focus on two-layer neural networks with ReLU activations and show that under a specific notion of separability in the RKHS induced by the infinite-width network, training (finitewidth) networks with stochastic gradient descent is robust against data poisoning attacks. Interestingly, we find that in addition to a lower bound on the width of the network, which is standard in the literature, we also require a distributiondependent upper bound on the width for robust generalization. We provide extensive empirical evaluations that support and validate our theoretical results.