MOMAT: Mixture of Multiple Atlases for Low-Power Jailbreak Defense of Quantized LLMs
MOMAT addresses jailbreak vulnerabilities in edge-deployed quantized LLMs by using a hardware-enhanced safety framework. It organizes safety data into multiple semantic atlases and uses Compute-in-Memory (CiM) for fast similarity search. This system provides domain-localized retrieval and intent detection, effectively mitigating the alignment weaknesses introduced by model quantization while maintaining low latency and high energy efficiency.