GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning

Researchers introduce GUDA, a machine unlearning-based method for attributing influence of training data groups to outputs in diffusion models. The approach approximates counterfactual scenarios without expensive full retraining, achieving ~100x speedup while more reliably identifying which artistic styles or object classes contributed to generated images compared to existing attribution methods.

GUDA addresses a fundamental challenge in generative AI transparency: understanding which training data influences model outputs at a group level rather than individual examples. This distinction matters because practitioners need to understand how broad categories—artistic movements, demographic representations, object classes—shape model behavior, not just trace individual images. The paper's core innovation replaces prohibitively expensive Leave-One-Group-Out retraining with machine unlearning applied to a pre-trained model, reducing computational costs dramatically while maintaining attribution accuracy.

The broader context reflects growing scrutiny of generative model training practices. As diffusion models like Stable Diffusion face legal challenges over copyright infringement and unauthorized use of artistic works, understanding exactly which training data influences outputs becomes legally and ethically important. Attribution methods provide accountability mechanisms and help mitigate risks of copyright disputes or harmful bias reproduction.

For the AI development community, GUDA's efficiency gains enable practical deployment of attribution analysis that was previously infeasible. Developers can now audit model behavior systematically without weeks of computational overhead. The reported ~100x speedup on CIFAR-10 and improved reliability over gradient-based methods suggest this approach could become standard practice for model evaluation and safety auditing.

Looking forward, similar unlearning-based attribution techniques may extend to other generative architectures and larger models. The practical viability demonstrated here could accelerate adoption of transparency measures in commercial AI products, potentially influencing regulatory expectations around model documentation and data provenance.

• →GUDA uses machine unlearning to approximate counterfactual training scenarios, eliminating expensive full-model retraining for group-level attribution.
• →The method achieves ~100x computational speedup compared to Leave-One-Group-Out retraining while improving attribution accuracy.
• →Testing on Stable Diffusion and CIFAR-10 demonstrates reliable identification of which training data groups (artistic styles, object classes) influence model outputs.
• →Group-level attribution has practical applications for copyright assessment, bias auditing, and regulatory compliance in generative AI systems.
• →Machine unlearning-based attribution may become standard practice for AI model transparency and safety evaluation.