Whose record is this? Diagnosing and authorizing record use in personalized multimodal models

The researchers proposed the concept of “record authorization,” requiring that the existence of the subject, the validity of record edges, and answer support all concurrently to determine whether records in personalized multimodal models can be used. To this end, a “RecordAuth-Diag” diagnostic suite containing 3,690 cases was constructed, and the model was tested by fixing queries and changing only images. Testing showed that Gemma-3-4B-IT had 63.69% local unauthorized usage at a clean recall rate of 25.75%, while CoViP had 26.02%. After implementing type-specific pre-generation authorization, Qwen’s card exposure rate on RecordAuth-Diag decreased from 43.63% to 3.06%, but the positive recall rate dropped from 86.26% to 60.90%. Full validation based on 560 DAVIS cases indicated that the improvement in security was mainly driven by support conditions rather than edge checks. The study conclusions only assess the effectiveness of contracts and do not address issues related to natural universality, consent, or visual identity.