Three researchers show that following the rate of iterations in a site's electricity consumption, at least 96% of model drives are detected without having to trust the operator.
Future AI governance agreements will need a way to verify that no large model training is taking place quietly on a monitored cluster, without depending on the good faith of its operator. Tom Kimpson, Mauricio Baker and Emlyn Graham propose to rely on the overall power consumption of the site. A synchronous drive produces a pulse whose cadence drifts slowly over the course of iterations; a Viterbi algorithm can track this signature.
Difficult and expensive to conceal
On simulated consumptions, when the operator does not try to hide, the method detects at least 96% of the training sessions in the face of a declared inference load, with 1% of false alarms, provided that the counter samples at least twice per second. Of the eight concealment strategies tested, only one thwarts the tool: to vary the amount of work very strongly from one iteration to the next. In its most advanced form, it hides training in up to 73% of five-minute windows, but at least 87% of these training sessions remain detected over a day.
This trick does not reduce throughput, but on small drives it requires spending 93% to 267% more computing to achieve the same performance. For the authors, a modest passive counter therefore makes the hidden drive both visible and expensive, making it a promising building block to check the use of computing power.
Read the full article on the original source
Go to the source →