Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes

TL;DR

This paper presents a geometric framework for analyzing memorization in diffusion models by measuring the sharpness of the log probability landscape, introduces new metrics for early detection, and proposes a mitigation strategy using sharpness-aware regularization.

Contribution

It introduces a novel geometric framework and metrics for memorization analysis in diffusion models, and develops a mitigation method based on sharpness-aware regularization.

Findings

The proposed metrics effectively quantify memorization sharpness.

Early-stage sharpness metrics provide insights into potential memorization.

Mitigation strategy reduces memorization by optimizing initial noise with regularization.

Abstract

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.