Diffusion Denoising as a Certified Defense against Clean-label Poisoning

2024-03-18

arXiv.org

AI Summary

This work addresses clean-label poisoning attacks—where only 1% of training samples contain imperceptible adversarial perturbations inducing targeted misclassification—by proposing the first provably robust defense leveraging pretrained diffusion models. Our method requires no model retraining: it purifies poisoned samples via diffusion-based denoising and integrates denoising smoothing with p-norm-bounded perturbation modeling to establish certified robustness guarantees. The key contribution is the first incorporation of off-the-shelf diffusion models into a provable defense framework for clean-label poisoning. Evaluated against seven state-of-the-art attacks, our approach reduces attack success rates to 0–16%, while incurring only marginal degradation in test accuracy. It thus achieves a favorable trade-off between certified robustness and model utility, substantially outperforming existing defenses.

Technology Category

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

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Large pretrained models with web data
• Security and Privacy: Data transparency and provenance

Abstract

We present a certified defense to clean-label poisoning attacks. These attacks work by injecting a small number of poisoning samples (e.g., 1%) that contain $p$-norm bounded adversarial perturbations into the training data to induce a targeted misclassification of a test-time input. Inspired by the adversarial robustness achieved by $denoised$ $smoothing$, we show how an off-the-shelf diffusion model can sanitize the tampered training data. We extensively test our defense against seven clean-label poisoning attacks and reduce their attack success to 0-16% with only a negligible drop in the test time accuracy. We compare our defense with existing countermeasures against clean-label poisoning, showing that the defense reduces the attack success the most and offers the best model utility. Our results highlight the need for future work on developing stronger clean-label attacks and using our certified yet practical defense as a strong baseline to evaluate these attacks.

Problem

• Defends against clean-label poisoning attacks using diffusion denoising
• Reduces attack success to 0-16% with minimal accuracy drop
• Provides certified robustness against adversarial training data tampering

Innovation

• Uses diffusion denoising for data sanitization
• Certified defense against clean-label poisoning
• Maintains high test accuracy post-defense