Towards Efficient and Domain-Agnostic Evasion Attack with High-dimensional Categorical Inputs

TL;DR

This paper introduces FEAT, a novel multi-armed bandit-based method for efficient, domain-agnostic adversarial attacks on classifiers with high-dimensional categorical inputs, addressing NP-hard challenges.

Contribution

The paper proposes FEAT, a new combinatorial search algorithm using UCB and OMP strategies, with theoretical guarantees and superior empirical performance.

Findings

FEAT outperforms existing attack methods in efficiency and effectiveness.

Theoretical bounds support FEAT's practical attack performance.

Empirical results demonstrate FEAT's applicability across diverse real-world datasets.

Abstract

Our work targets at searching feasible adversarial perturbation to attack a classifier with high-dimensional categorical inputs in a domain-agnostic setting. This is intrinsically an NP-hard knapsack problem where the exploration space becomes explosively larger as the feature dimension increases. Without the help of domain knowledge, solving this problem via heuristic method, such as Branch-and-Bound, suffers from exponential complexity, yet can bring arbitrarily bad attack results. We address the challenge via the lens of multi-armed bandit based combinatorial search. Our proposed method, namely FEAT, treats modifying each categorical feature as pulling an arm in multi-armed bandit programming. Our objective is to achieve highly efficient and effective attack using an Orthogonal Matching Pursuit (OMP)-enhanced Upper Confidence Bound (UCB) exploration strategy. Our theoretical analysis bounding the regret gap of FEAT guarantees its practical attack performance. In empirical analysis, we compare FEAT with other state-of-the-art domain-agnostic attack methods over various real-world categorical data sets of different applications. Substantial experimental observations confirm the expected efficiency and attack effectiveness of FEAT applied in different application scenarios. Our work further hints the applicability of FEAT for assessing the adversarial vulnerability of classification systems with high-dimensional categorical inputs.