Dormant key: Unlocking universal adversarial control in text-to-image models.

Abstract

Text-to-Image (T2I) diffusion models have gained significant traction due to their remarkable image generation capabilities, raising growing concerns over the security risks associated with their use. Prior studies have shown that malicious users can subtly modify prompts to produce visually misleading or Not-Safe-For-Work (NSFW) content, even bypassing existing safety filters. Existing adversarial attacks are often optimized for specific prompts, limiting their generalizability, and their text-space perturbations are easily detectable by current defenses. To address these limitations, we propose a universal adversarial attack framework called dormant key. It appends a transferable suffix that can be appended as a "plug-in" to any text input to guide the generated image toward a specific target. To ensure robustness across diverse prompts, we introduce a novel hierarchical gradient aggregation strategy that stabilizes optimization over prompt batches. This enables efficient learning of universal perturbations in the text space, improving both attack transferability and imperceptibility. Experimental results show that our method effectively balances attack performance and stealth. In NSFW generation tasks, it bypasses major safety mechanisms, including keyword filtering, semantic analysis, and text classifiers, and achieves over 18 % improvement in success rate over baselines.