Training data attribution in diffusion models via mirrored unlearning and noise-consistent skew

2026-05-18

AI Summary

Current methods for attributing training data in diffusion models lack reliability and robustness, limiting their applicability in real-world scenarios. This work proposes the Mirror Unlearning with Consistency Skew (MUCS) framework, which achieves efficient attribution by performing bounded mirror gradient ascent fine-tuning on the original model and computing a normalized skew between the outputs of the original and modified models on consistent noise samples. Conceptually simple and broadly applicable, MUCS significantly outperforms state-of-the-art approaches across three standard benchmarks, demonstrating strong reliability, robustness, and generalization. Furthermore, the method reveals intrinsic instance overlap among generated samples and highlights the potential for ensemble strategies in attribution techniques.

Abstract

Training data attribution (TDA) should enable generative model interpretability and foster a variety of related downstream tasks. Nonetheless, current TDA approaches lack reliability and robustness, preventing their adoption in real-world setups. In this paper, we take a decisive step towards more reliable and robust TDA for diffusion models. We propose to perform TDA with mirrored unlearning and noise-consistent skew (MUCS). The idea is to fine-tune a second model with bounded mirrored gradient ascent, and to measure the normalized skew of this model with respect to the original one using consistent noise samples. We show that, while being conceptually simple and generic, MUCS systematically outperforms existing methods on three different datasets by a large margin. We additionally study the effect that core design choices have on final performance, and analyze novel aspects regarding the overlap of influential instances across generated items and the potential of ensembling TDA approaches. We believe that our findings may have broader implications for more general unlearning setups, as well as for tasks requiring the comparison of diffusion losses.

Problem

• training data attribution
• diffusion models
• reliability
• robustness
• generative model interpretability

Innovation

• training data attribution
• diffusion models
• mirrored unlearning
• noise-consistent skew
• model interpretability