MUCS: New Method Improves Training Data Attribution for Diffusion Models

Researchers have proposed MUCS (Mirrored Unlearning and Noise-Consistent Skew), a new approach to training data attribution (TDA) for diffusion models, addressing longstanding reliability and robustness gaps in existing methods. The technique fine-tunes a secondary model using bounded mirrored gradient ascent, then measures the normalized skew between that model and the original using consistent noise samples. The authors report that MUCS systematically outperforms current TDA methods across three datasets by a significant margin.

Beyond attribution accuracy, the paper examines how influential training instances overlap across generated outputs and explores the potential of ensembling multiple TDA approaches. The authors suggest the findings may carry broader implications for machine unlearning and for any task requiring comparison of diffusion model losses — areas of growing relevance as the media and creative industries face mounting pressure around copyright, data provenance, and model interpretability.