FedAttr enables privacy-preserving client-level attribution of watermarked data in federated LLM fine-tuning via paired-subset differencing, differential scoring, and Stouffer combination, achieving 100% TPR and 0% FPR with bounded leakage.
Watermarking Makes Language Models Radioactive , booktitle =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
FedAttr: Towards Privacy-preserving Client-Level Attribution in Federated LLM Fine-tuning
FedAttr enables privacy-preserving client-level attribution of watermarked data in federated LLM fine-tuning via paired-subset differencing, differential scoring, and Stouffer combination, achieving 100% TPR and 0% FPR with bounded leakage.