New Probabilistic Framework for Training Data Attribution in Generative AI

A new arXiv preprint (2607.21615) introduces FrED, a probabilistic framework for Training Data Attribution that operates in a black-box setting without requiring access to model weights. The method fuses continuous feature similarities with domain-specific Knowledge Graphs (KGs) to ground attribution in structural reality, rewarding specific historical samples while suppressing generic background data. The framework was evaluated across two domains: abstract artistic image synthesis and high-dimensional physical weather forecasting. Benchmarking shows robust performance in linking outputs to data and domain context. The work addresses critical needs for transparency and accountability in generative AI deployment.