Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·11d ago·5 min read
Learning Normal Diffusion Dynamics for Backdoor Defense in Text-to-Image Models
TL;DR
Normal Diffusion Dynamics Learning (NDDL) is a new framework designed to defend text-to-image diffusion models against backdoor attacks by learning normal transition patterns from benign samples. It effectively detects deviations indicative of backdoor presence and localizes triggers without prior knowledge of the attack.
✦ Why It Matters
Implement NDDL to enhance the security of your text-to-image models against evolving backdoor threats.
Key Takeaways
- Adopt NDDL for robust backdoor defense in T2I models.
- Utilize only benign samples to train your defense model.
- Leverage predicted transition dynamics to identify anomalies.
- Implement low-semantic substitutions for effective trigger localization.
How It Works
NDDL learns the normal transition dynamics of diffusion trajectories by analyzing benign samples. It constructs compact representations across multiple spaces and trains a timestep-conditioned model to predict how diffusion should evolve.
During inference, it compares observed transitions against these predictions to detect deviations that indicate backdoor attacks.