betaFeedback

WorksMathematics

Nonparametric Classification under Adversarial Attacks: Decision-Boundary Geometry and Optimal Rates

Abstract

Unlike in continuous regression, a small input perturbation in classification can move an input across a decision boundary and flip the predicted label. Decision-boundary geometry is therefore essential to understanding adversarial classification risk. Under anisotropic smoothness and margin conditions, we quantify this geometric contribution and establish sharp minimax rates for adversarial excess risk relative to the ordinary Bayes risk. These rates comprise a statistical term governed by the harmonic mean smoothness and a geometric cost of adversarial perturbations governed by the least smooth direction. As smoothness and the margin exponent vary, the adversarial term exhibits a phase transition from a fractional power of the attack radius to linear scaling, with a logarithmic correction in a critical regime. In addition, we construct an adaptive classifier that attains these rates over a range of smoothness and margin classes. We further characterize the minimax rate for adversarial excess risk under a budget on ordinary excess risk, yielding an explicit trade-off curve. The curve identifies attack regimes where ordinary minimax optimality entails a suboptimal adversarial rate. Finally, numerical experiments support the theoretical findings.

Paper details

vlnc:W_3nd8746x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…