LLM-Attacker Framework for Split Model Privacy Evaluation
October 14, 2026
LLM-Attacker is a unified red-teaming framework designed to evaluate privacy risks in split and distributed LLM architectures. In GPT-2 experiments, the framework demonstrated that including training gradients increased token recovery from 94.20% to 97.38%.
HOW THIS AFFECTS YOU
●
researcherYou can use this to benchmark defenses against intermediate representation reconstruction.
●
policyThis highlights critical privacy vulnerabilities in distributed inference architectures.