Attribute Inference Attack on Edge-Cloud Split Multimodal LLM Features

Mon, 26 Oct 2026 18:39:19 +0000

Engineer — Learn: If you’re designing or auditing a split-inference LLM pipeline, this research challenges the assumption that keeping raw images on-device is sufficient privacy protection — intermediate features still leak age, gender, and race. No patch exists; the takeaway is an architectural one: evaluate representation-level privacy, not just response-level filtering, before deploying such pipelines.
SOC/IR — Skip
Leader — Learn: Relevant for AI governance: if your organization deploys split-inference LLMs that process biometric attributes, this research demonstrates that edge-local raw data does not prevent cloud-side inference of protected characteristics — a GDPR/HIPAA consideration worth flagging to your AI/privacy counsel when reviewing new deployments.

Split-Inference on CuraSec