Evasion attacks on generative safeguards: target-concept reappearance under structural control in grey-box settings
Open Access
- 8 October 2026
- journal article
- research article
- Published by Springer Nature in Cybersecurity
- Vol. 9 (1) , 227
Abstract
Concept erasure has emerged as a practical safeguard for suppressing unsafe or unwanted concepts in text-to-image diffusion models. However, existing methods are primarily designed for text-triggered generation and are usually evaluated under text-only prompting. In this paper, we identify a practical robustness gap by showing that concepts suppressed under text-only evaluation can reappear after a concept-erased backbone is composed with a compatible structural controller and supplied with target-consistent, information-bearing structural conditions. We propose Structure-Guided Concept Reappearance (SGCR), an inference-time attack that combines a benign prompt with a structural condition, such as a Canny edge map, depth map, or segmentation map, to induce target-related outputs without retraining, gradient-based optimization, or modification of the defended backbone. Extensive experiments on explicit content, copyrighted artistic styles, and generic objects demonstrate that SGCR remains effective across multiple representative erasure defenses and structural modalities. These findings show that suppression established under text-only evaluation may not persist in controller-augmented diffusion pipelines. The demonstrated threat applies to local modular systems and services exposing compatible structural-control interfaces, rather than closed text-only APIs.Keywords
Funding Information
- National Natural Science Foundation of China (62506339)
- National Natural Science Foundation of China (62476250)
- National Natural Science Foundation of China (62472335)
- National Natural Science Foundation of China (62372137)
- Zhejiang Provincial Natural Science Foundation of China (LQN25F010018)
- Science and Technology Plan of Zhejiang Province (2025C01SA100283)
- Specialized Foundation of Zhejiang University of Science and Technology (2025QN060)
- Postgraduate Research and Innovation Foundation of Zhejiang University of Science and Technology (2024yjskc19)
This publication has 29 references indexed in Scilit:
- R.A.C.E. : Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model2024
- MACE: Mass Concept Erasure in Diffusion Models2024
- FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition2024
- A Survey of Multimodal Controllable Diffusion ModelsJournal of Computer Science and Technology, 2024
- Transferable adversarial sample purification by expanding the purification space of diffusion modelsThe Visual Computer, 2024
- Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models2023
- Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation2023
- A survey on adversarial attacks and defenses for object detection and their applications in autonomous vehiclesThe Visual Computer, 2022
- High-Resolution Image Synthesis with Latent Diffusion Models2022
- The Unreasonable Effectiveness of Deep Features as a Perceptual Metric2018