Adversarial Pruning: A Survey and Benchmark of Pruning Methods for Adversarial Robustness

2024-09-02

Pattern Recognition

AI Summary

Existing adversarial robustness-aware pruning methods lack standardized evaluation protocols, hindering fair and systematic comparison. Method: We introduce the first standardized benchmarking framework for robustness-oriented pruning, conducting comprehensive empirical analysis of 32 state-of-the-art methods—including structured/unstructured, sensitivity/gradient-based, and joint robust training paradigms—across CIFAR-10, CIFAR-100, and Tiny-ImageNet. Contribution/Results: We identify a non-monotonic relationship between sparsity and adversarial robustness, leading to a novel robustness-aware pruning taxonomy. Experimental results show that only five methods simultaneously preserve clean accuracy and maintain robustness under PGD and AutoAttack. This work establishes a reproducible benchmark and provides theoretical insights for secure, compressed AI models.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Security and Privacy: Large-scale security measurements

Problem

• Survey and categorize adversarial pruning methods for robustness
• Propose fair benchmark to evaluate pruning methods effectively
• Empirically analyze top-performing adversarial pruning traits

Innovation

• Survey and taxonomy of adversarial pruning methods
• Novel fair evaluation benchmark proposal
• Empirical re-evaluation of pruning techniques