Deeper Diffusion Models Amplify Bias

TL;DR

This paper investigates how deep diffusion models can amplify biases present in training data and potentially compromise privacy, highlighting risks associated with their internal workings and the bias-variance tradeoff.

Contribution

It provides a systematic theoretical and empirical analysis of bias amplification and privacy risks in deep diffusion models, expanding understanding beyond traditional generalization.

Findings

Deeper diffusion models tend to amplify training data bias.

Diffusion models can compromise training sample privacy.

Theoretical and empirical validation of bias and privacy risks.

Abstract

Despite the remarkable performance of generative Diffusion Models (DMs), their internal working is still not well understood, which is potentially problematic. This paper focuses on exploring the important notion of bias-variance tradeoff in diffusion models. Providing a systematic foundation for this exploration, it establishes that at one extreme, the diffusion models may amplify the inherent bias in the training data, and on the other, they may compromise the presumed privacy of the training samples. Our exploration aligns with the memorization-generalization understanding of the generative models, but it also expands further along this spectrum beyond "generalization", revealing the risk of bias amplification in deeper models. Our claims are validated both theoretically and empirically.