Membership Inference for Contrastive Pre-training Models with Text-only PII Queries

Researchers developed UMID, a new text-only auditing framework to detect if personally identifiable information was memorized during training of multimodal AI models like CLIP and CLAP. The method significantly improves efficiency and effectiveness of membership inference attacks while maintaining privacy constraints.

• →UMID enables privacy auditing of large multimodal models using only text queries, avoiding direct exposure of sensitive biometric data.
• →The framework performs text-guided cross-modal latent inversion to extract similarity and variability signals for detection.
• →UMID delivers strong detection performance with sub-second auditing costs, making it practically viable for large-scale models.
• →Traditional shadow-model membership inference attacks are computationally prohibitive for large multimodal backbones.
• →The research addresses growing concerns about web-scale trained models memorizing personally identifiable information.