Early Detection of Re-Identification Risk in Multi-Turn Dialogues via Entity-Aware Evidence Accumulation

초록

In multi-turn conversational AI, individually innocuous personally identifiable information (PII) fragments disclosed across successive turns can accumulate into a re-identification risk that no single utterance reveals on its own. Existing PII detectors operate on isolated utterances and therefore cannot track this cross-turn evidence build-up. We propose a stateful middleware guardrail whose core design principle is speaker-attributed entity isolation: every extracted PII fragment is attributed to its originating conversational participant, and evidence is accumulated in entity-isolated subgraphs that prevent cross-entity contamination. The system signals re-identification onset tpred at the earliest turn where combination-based rules grounded in the uniqueness literature are satisfied. On a 184-record template-synthetic evaluation corpus, the gated NER configuration leads on primary timeliness (OW@5 = 73.4%, MAE= 1.357 turns); the full system achieves OW@5 = 70.7% with MAE = 2.442 turns as an alternative operating mode for ambiguity-sensitive disclosure patterns. We further evaluate behavior on a 300-record mutation stress set, test RULE_B on the ABCD external corpus, and supplement RULE_A evaluation with both a proxy-labeled transfer analysis on PersonaChat and a manual annotation study on 151 Switchboard dialogues. The reported results should be interpreted as an initial empirical reference point rather than a sufficient endpoint for autonomous runtime enforcement.

키워드