Protecting Multimodal Datasets from Unauthorized Use
A novel fingerprinting approach for vision-language models
PATFinger introduces a transferable fingerprinting method that protects multimodal datasets against unauthorized usage without compromising model performance.
- Non-intrusive technique that verifies dataset ownership across modalities
- Uses prompt-adapted fingerprinting to embed verifiable patterns
- Maintains model accuracy while providing robust security measures
- Addresses the growing concern of dataset misuse in AI development
This research provides critical security infrastructure for organizations investing in multimodal datasets, helping to protect intellectual property as vision-language models become more prevalent in business applications.