Research

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

99.0% classification accuracy across 10 open-weight models, distinguishing model architectures and quantization states from non-intrusive hardware traces.