Reproduce “Watch Your Step: Information Injection in Diffusion Models via Shadow Timestep Embedding” on OpenResearch

About this paper

An Huang, Junggab Son, Zuobin Xiong

Researchers at the University of Nevada Las Vegas introduced Shadow Timestep Embedding (STE), a method to inject and retrieve information covertly within diffusion models by exploiting underutilized temporal embedding spaces. This approach demonstrated dual-use capabilities, enabling highly stealthy backdoor attacks with 99.2% attack success rate on CIFAR-10 and robust watermark verification comparable to existing state-of-the-art techniques.

Key insights

• The timestep embedding component in diffusion models possesses substantial, underutilized informational capacity that can be leveraged to encode independent data distributions.
• Different timestep intervals, particularly those generated by sufficiently large shadow offsets, exhibit low mutual coherence, ensuring that learned information in one interval does not interfere with others.
• The diffusion model's temporal embedding space functions as a previously overlooked side channel, enabling both stealthy information injection for attacks and robust, temporally disentangled watermarking for defense.

Summary by alphaXiv.