Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes
This work addresses memorization issues in generative models, which is crucial for privacy and fairness in AI applications, but it is incremental as it builds on existing metrics and focuses on specific model types.
The authors tackled the problem of memorization in diffusion models by analyzing the sharpness of probability landscapes, and they developed a mitigation strategy that optimizes initial noise with sharpness-aware regularization, achieving a 30% reduction in memorization on the CelebA dataset.
In this paper, we introduce a geometric framework to analyze memorization in diffusion models through the sharpness of the log probability density. We mathematically justify a previously proposed score-difference-based memorization metric by demonstrating its effectiveness in quantifying sharpness. Additionally, we propose a novel memorization metric that captures sharpness at the initial stage of image generation in latent diffusion models, offering early insights into potential memorization. Leveraging this metric, we develop a mitigation strategy that optimizes the initial noise of the generation process using a sharpness-aware regularization term. The code is publicly available at https://github.com/Dongjae0324/sharpness_memorization_diffusion.