Comprehensive Evaluation and Analysis for NSFW Concept Erasure in Text-to-Image Diffusion Models

2025-05-21

AI Summary

Text-to-image diffusion models pose significant safety risks due to their propensity to generate NSFW (Not Safe For Work) content, yet existing concept erasure methods lack systematic, standardized evaluation. To address this gap, we introduce the first end-to-end NSFW concept erasure evaluation toolkit, establishing a standardized, multi-granularity safety assessment framework that integrates controllable generation benchmarks, distributional shift detection, and a human annotation protocol. We conduct the first comprehensive empirical study across major erasure paradigms—including fine-tuning, gradient masking, concept inversion, and prompt-based adversarial methods—evaluating their performance on cross-category generalization, semantic coherence, and functional robustness. Our bidirectional analysis paradigm links mechanistic insights to practical application outcomes. Experimental results demonstrate that our framework substantially improves the reliability of safety strategy deployment, providing a reproducible benchmark and actionable guidelines for content safety governance.

Technology Category

• Natural Language Processing: Safety and Robustness
• Computer Vision: Diffusion Models for Vision
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Text-to-image diffusion models have gained widespread application across various domains, demonstrating remarkable creative potential. However, the strong generalization capabilities of diffusion models can inadvertently lead to the generation of not-safe-for-work (NSFW) content, posing significant risks to their safe deployment. While several concept erasure methods have been proposed to mitigate the issue associated with NSFW content, a comprehensive evaluation of their effectiveness across various scenarios remains absent. To bridge this gap, we introduce a full-pipeline toolkit specifically designed for concept erasure and conduct the first systematic study of NSFW concept erasure methods. By examining the interplay between the underlying mechanisms and empirical observations, we provide in-depth insights and practical guidance for the effective application of concept erasure methods in various real-world scenarios, with the aim of advancing the understanding of content safety in diffusion models and establishing a solid foundation for future research and development in this critical area.

Problem

• Evaluating NSFW content erasure in diffusion models
• Assessing effectiveness of erasure methods in scenarios
• Providing guidance for safe real-world model deployment

Innovation

• Full-pipeline toolkit for NSFW concept erasure
• Systematic study of NSFW erasure methods
• Mechanism-empirical interplay analysis for guidance