Adaptive versus Standard Descent Methods and Robustness Against Adversarial Examples

Abstract

Adversarial examples are a pervasive phenomenon of machine learning models where seemingly imperceptible perturbations to the input lead to misclassifications for otherwise statistically accurate models. In this paper we study how the choice of optimization algorithm influences the robustness of the resulting classifier to adversarial examples. Specifically we show an example of a learning problem for which the solution found by adaptive optimization algorithms exhibits qualitatively worse robustness properties against both $L_2$- and $L_{\infty }$-adversaries than the solution found by non-adaptive algorithms. Then we fully characterize the geometry of the loss landscape of $L_2$-adversarial training in least-squares linear regression. The geometry of the loss landscape is subtle and has important consequences for optimization algorithms. Finally we provide experimental evidence which suggests that non-adaptive methods consistently produce more robust models than adaptive methods.

References in corpus (9)

• Theoretically Principled Trade-off between Robustness and Accuracy
• Provable defenses against adversarial examples via the convex outer adversarial polytope
• The Marginal Value of Adaptive Gradient Methods in Machine Learning
• Adversarially Robust Generalization Requires More Data
• Poincaré GloVe: Hyperbolic Word Embeddings
• On the Geometry of Adversarial Examples
• Cross-Entropy Loss and Low-Rank Features Have Responsibility for Adversarial Examples
• SGD Converges to Global Minimum in Deep Learning via Star-convex Path
• Adversarial Training with Voronoi Constraints