Privacy-Preserving Text Sanitization for Distributed Agents Collaboration via Disentangled Representations

In three linesDiSan, a privacy-preserving sanitization framework, factorizes text into two subspaces: one preserving task semantics and one containing stylistic signatures. On a distributed multi-agent RAG benchmark, DiSan reduces PII exposure by 20× while maintaining 83% answer faithfulness, and lowers Enron stylometric attribution by 73.2% (TF-IDF) and 70.6% (neural probe).

Summary generated by Claude — human-verified