Classifier-Free Guidance inside the Attraction Basin May Cause Memorization

2024-11-23

arXiv.org

AI Summary

Diffusion models are prone to exact memorization of training images during denoising due to the “basin-of-attraction” phenomenon, posing significant copyright and privacy risks. This work introduces, for the first time, a basin-of-attraction theory to characterize the dynamical mechanism underlying memorization. Building on trajectory-level analysis of the denoising process, we propose two novel guidance paradigms—delayed classifier-free guidance (CFG) and reverse guidance—that actively steer sampling paths away from memorized attractors, enabling non-memorizing generation. Our approach requires no model retraining or dataset curation and is fully compatible with existing diffusion frameworks. Extensive experiments across diverse benchmarks empirically validate the existence of memorization basins and demonstrate substantial reductions in memorization rates while preserving image fidelity and text–image alignment. The core contributions are: (1) a dynamical attribution model that formally links memorization to stochastic trajectory convergence; and (2) the first guidance scheduling mechanism explicitly designed to suppress basin attraction.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Neural Architectures and Foundation Models
• Natural Language Processing: Generation

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Diffusion models are prone to exactly reproduce images from the training data. This exact reproduction of the training data is concerning as it can lead to copyright infringement and/or leakage of privacy-sensitive information. In this paper, we present a novel way to understand the memorization phenomenon, and propose a simple yet effective approach to mitigate it. We argue that memorization occurs because of an attraction basin in the denoising process which steers the diffusion trajectory towards a memorized image. However, this can be mitigated by guiding the diffusion trajectory away from the attraction basin by not applying classifier-free guidance until an ideal transition point occurs from which classifier-free guidance is applied. This leads to the generation of non-memorized images that are high in image quality and well-aligned with the conditioning mechanism. To further improve on this, we present a new guidance technique, emph{opposite guidance}, that escapes the attraction basin sooner in the denoising process. We demonstrate the existence of attraction basins in various scenarios in which memorization occurs, and we show that our proposed approach successfully mitigates memorization.

Problem

• Mitigate memorization in diffusion models to prevent copyright and privacy issues.
• Identify attraction basins causing exact reproduction of training data images.
• Propose new guidance techniques to generate high-quality, non-memorized images.

Innovation

• Mitigate memorization via delayed classifier-free guidance
• Introduce opposite guidance to escape attraction basins
• Generate high-quality, non-memorized, condition-aligned images