Research | Can power consumption reveal whether a data center is secretly training AI?

During training, compute-intensive phases alternate with periods of communication between graphics processing units (GPUs). This causes power consumption to fluctuate in a characteristic rhythm, roughly once per second. The rhythm is not perfectly constant, but changes with workload and hardware state.

The proposed method tracks the rhythm even when its frequency changes. To do so, the authors use a so-called Viterbi tracker, a technique for detecting temporal patterns that is also used in fields such as signal processing.

The sensor is located outside the data center operator's control. It therefore does not need to trust the operator's internal logs or statements about the computations being performed. This makes the method independent of the operator.

The authors compare the approach to a scenario in which an operator runs AI models as usual while secretly training an additional model. The goal is not simply to detect high power consumption, but to identify the specific temporal pattern left by the training process. This makes it possible to isolate and detect training activity even when multiple AI models are running in parallel.

In simulated data center workloads, the method detected at least 96 percent of training runs, provided the operator did not take deliberate steps to conceal them. The researchers also tested eight concealment strategies, which are costly and time-consuming to implement. The false positive rate was 1 percent.

The study addresses a fundamental challenge in AI regulation: how can you verify what a company is actually computing when you cannot trust its own statements? Power measurements capture a physical side effect of computation without requiring access to the operator's software. The full study is available at the link below.