The authors simulate datacentre power traces that mix training and inference workloads, passed through a model of a facility-level meter (low-pass filtered, 2 Hz or slower).
Viterbi tracker detects hidden AI training from facility power
The method catches at least 96% of training runs at a 1% false-alarm rate with a 2 Hz meter, and strongest hiding costs 93–267% more compute.
Top University
Tom Kimpson · Mauricio Baker · Emlyn Graham
The University of Melbourne · University of Oxford · Machine Alignment, Transparency & Security (MATS)
Research Digest··3 min read
The authors model datacentre power and develop a Viterbi-based detector that follows the wandering iteration cadence of synchronous training.
§