Non-Intrusive Characterization of Local Large Language Models via Hardware Side Channels

Can local Large Language Model deployments be observed and audited without modifying model binaries, weights, or host environments? This project investigates whether physical hardware side-channels—specifically NVML GPU telemetry and CPU cache access timing—can serve as non-intrusive observability mechanisms for local inference workloads.

Built a C/Python acquisition and alignment pipeline integrating NVML GPU telemetry and Flush+Reload cache timing into PyTorch tensors. Developed a dual-branch 1D CNN with late fusion for model architecture and quantization-state classification.

Achieved 99.0% classification accuracy across 10 open-weight model and quantization configurations, demonstrating that the measured non-intrusive hardware traces contain sufficient signal to distinguish architecture and quantization state.