Abstract Diffusion models (DMs) have emerged as a dominant model for high-fidelity generative tasks across vision, text, and audio domains. However, their increasing deployment raises serious concerns about backdoor vulnerabilities. Existing attacks primarily rely on static and perceptible triggers, making them susceptible to detection via state-of-the-art inversion-based defenses. In this paper, we propose MACTDiffusion, a novel Multi-Agent Backdoor Attack framework for diffusion models empowered by Chaotic Trigger Controller (CTC). Unlike conventional single-trigger attacks, our method employs multiple imperceptible sub-triggers that are independently ineffective but collectively activate the backdoor when spatiotemporally aligned. To further enhance stealth and unpredictability, we leverage chaotic maps to randomly control the trigger injection time, spatial location, and active sub-trigger combinations during both training and inference. This makes the attack behave unpredictably, avoids leaving detectable patterns, and significantly increases the difficulty for defenders to analyze or reconstruct the trigger mechanism. Furthermore, the proposed approach is compared with leading defenses such as Elijah and TERD. This study reveals a novel type of stealthy backdoor threat that has not been explored before, highlighting the urgent need for stronger and more reliable security measures for generative models.
Bikkina et al. (Thu,) studied this question.