Detecting Hidden ML Training With Zero-Overhead Telemetry

2026-06-17

AI Summary

Existing GPU telemetry–based approaches for monitoring AI computation are vulnerable to evasion and lack adversarial robustness. This work proposes a zero-overhead, privacy-preserving detection framework that relies solely on NVML telemetry signals to identify covert machine learning training tasks by analyzing the physical effects they induce on GPU hardware, without requiring access to model weights or training data. We present the first systematic modeling of adversarial evasion strategies and rigorously evaluate detector robustness through multiple rounds of adversarial iterations. Experimental results demonstrate that the proposed method achieves 98.2% binary classification accuracy across the full dataset and maintains detection accuracy between 43% and 87% against previously unseen adversarial camouflage workloads, substantially enhancing the evasion resistance of AI regulatory systems.

Abstract

Hardware-enabled monitoring of GPU workloads underpins many proposals for AI compute governance, but if developers can defeat monitoring mechanisms, such schemes are unworkable. We evaluate the adversarial robustness of GPU workload classification using only zero-overhead, privacy-preserving NVML telemetry: content-agnostic signals that observe physical effects of computation without accessing model weights, training data, or hyperparameters. Across 5 rounds of monitor-evader iteration, we evaluate 20 evasion strategy families on 9 GPU models spanning 4 architecture generations. We develop a classifier that achieves 98.2% binary accuracy at identifying training workloads across the whole corpus, and 43-87% accuracy against the most challenging unexpected workloads even when they are adversarially disguised.

Problem

• AI governance
• GPU telemetry
• adversarial evasion
• ML training detection
• zero-overhead monitoring

Innovation

• zero-overhead telemetry
• adversarial robustness
• GPU workload classification
• privacy-preserving monitoring
• AI governance