Evidence · DiggingBeagle record
Research on sensitive attribute recovery from split multimodal model activations
Research on sensitive attribute recovery from split multimodal model activations
- Source role
- primary disclosure
Evidence record
Research on sensitive attribute recovery from split multimodal model activations
Claim-level citations (2)
- supportsMultimodal activation sharing can leak attributes hidden from model outputs: The split-model study reports that with Safe-LLaVA response protections, 20 labeled examples per sensitive-attribute class still recover approximately 75 to 85 percent of the leakage measured on an unprotected intermediate representation.
Abstract, Safe-LLaVA and 20 labeled samples per class
- supportsMultimodal activation sharing can leak attributes hidden from model outputs: The paper reports sensitive-attribute inference from transmitted intermediate multimodal representations with limited labeled examples, recovering a substantial share of the reference leakage despite output-facing protections.
Abstract; split edge/server threat model; 20-example attack configuration