Evaluating Adversarial Protections for Diffusion Personalization: A Comprehensive Study

TL;DR

This paper systematically compares eight perturbation-based privacy protection methods for diffusion models across different domains, providing practical insights into their effectiveness and imperceptibility under various conditions.

Contribution

It offers a comprehensive evaluation of multiple protection techniques for diffusion models, highlighting their strengths and limitations in real-world scenarios.

Findings

AdvDM and MetaCloak show high protection efficacy.

FSGM and SDS maintain better visual imperceptibility.

Protection performance varies with perturbation budgets and domain types.

Abstract

With the increasing adoption of diffusion models for image generation and personalization, concerns regarding privacy breaches and content misuse have become more pressing. In this study, we conduct a comprehensive comparison of eight perturbation based protection methods: AdvDM, ASPL, FSGM, MetaCloak, Mist, PhotoGuard, SDS, and SimAC--across both portrait and artwork domains. These methods are evaluated under varying perturbation budgets, using a range of metrics to assess visual imperceptibility and protective efficacy. Our results offer practical guidance for method selection. Code is available at: https://github.com/vkeilo/DiffAdvPerturbationBench.