Carrier Current Sensing for Computation-Level Bias Detection in Large Language Models: A Methodological Manual

Abstract

Auditing large language models (LLMs) for demographic bias typically relies on analyzing text outputs, which cannot detect cases where a model computes differently across demographic inputs yet produces equivalent output — a phenomenon we term dark bias. We investigate whether such computation-level differences are detectable through carrier current analysis: non-invasive monitoring of a GPU's power-cable current during inference. Building on prior work showing that power and electromagnetic side channels can recover neural network architecture and workload characteristics, we test the novel hypothesis that current traces also carry information about the demographic content of an input prompt. Using counterfactual prompt pairs that vary only a name drawn from validated demographic name sets, under deterministic decoding and token-count-matched conditions, we record current traces via a Hall-effect clamp and classify demographic group from extracted trace features using cross-validated logistic regression. Critically, we assess whether any detected signal persists when text outputs are matched across groups (output-controlled analysis) — the key test that would distinguish computation-level bias from bias already visible at the output. We report classification performance with bootstrapped confidence intervals rather than point estimates, given the small expected effect size and the possibility that the underlying signal does not exist at this level of measurement. A positive result would establish current-based side-channel analysis as a novel bias-detection method; a null result would meaningfully constrain where such bias, if present, is more likely to reside.