arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2509.25792cs.AIcs.CV

PureVQ-GAN:通过向量量化瓶颈防御数据投毒攻击

PUREVQ-GAN: Defending Data Poisoning Attacks through Vector-Quantized Bottlenecks

University of California, Los Angeles(加州大学洛杉矶分校);California Institute of Technology(加州理工学院)

Alexander Branch, Omead Pooladzandi, Radin Khosraviani, Sunay Gajanan Bhat, Jeffrey Jiang, Gregory Pottie

AI总结:

提出 PureVQ-GAN,通过向量量化离散瓶颈和 GAN 判别器消除后门触发器,在 CIFAR-10 上以极低投毒成功率、高干净准确率和超过 50 倍加速实现数据投毒防御。

摘要:

我们提出 PureVQ-GAN,这是一种针对数据投毒的防御方法,它利用带 GAN 判别器的向量量化 VAE,迫使后门触发器经过离散瓶颈。通过使用学习到的码本对投毒图像进行量化,PureVQ-GAN 在保留语义内容的同时破坏细粒度触发器模式。GAN 判别器确保输出符合自然图像分布,防止重建分布外扰动。在 CIFAR-10 上,PureVQ-GAN 对 Gradient Matching 和 Bullseye Polytope 攻击实现 0% 投毒成功率(PSR),对 Narcissus 攻击实现 1.64% 的投毒成功率,同时保持 91–95% 的干净准确率。与需要数百次迭代细化步骤的基于扩散的防御方法不同,PureVQ-GAN 的速度提升超过 50 倍,使其适用于实际训练流程。

英文摘要:

We introduce PureVQ-GAN, a defense against data poisoning that forces backdoor triggers through a discrete bottleneck using Vector-Quantized VAE with GAN discriminator. By quantizing poisoned images through a learned codebook, PureVQ-GAN destroys fine-grained trigger patterns while preserving semantic content. A GAN discriminator ensures outputs match the natural image distribution, preventing reconstruction of out-of-distribution perturbations. On CIFAR-10, PureVQ-GAN achieves 0% poison success rate (PSR) against Gradient Matching and Bullseye Polytope attacks, and 1.64% against Narcissus while maintaining 91-95% clean accuracy. Unlike diffusion-based defenses requiring hundreds of iterative refinement steps, PureVQ-GAN is over 50x faster, making it practical for real training pipelines.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑