Project description

What this project is

My previous research project developed a workload classifier that distinguishes training, inference, and non-ML computations, based on short snapshots of NVML telemetry logs (which record utilization of different GPU components). However, this system has security vulnerabilities, both in NVIDIA software libraries, and inherently in any operating system that is running alongside the ML workload, which could allow an evader to spoof the telemetry and insert readings that cause a training workload to be misclassified as inference or something else.

Project is planned by me, and physical construction is done by my research fellow Felix. When all the hardware is running, I'll do the ML experiments in my existing project repo, plus new ones. For context, see my paper <https://arxiv.org/abs/2606.19262> and repository <https://github.com/robirahman/GPU-monitoring/>

Theory of impact

This makes AI governance more reliable and prevents developers from cheating by disguising training workloads as inference or non-ML computations.

How the money will be spent

Computing hardware:
6x RTX 3090 @ $900 each
16x DDR4 32GB @ $75
2x cooler @ $1000
2x GPU rig frame with 8x PCIe risers + fans @ $200
4x 1500W PSU @ $200
2x Add2PSU sync adapter @ $15
2x 2TB SSD @ $140
8x 16TB HDD @ $140

Sensors:
Raspberry Pi 5 8GB @ $120
FLIR Lepton 3.5 thermal camera @ $400
6x microphones @ $10
12x temperature sensors @ $2
12x Hall effect magnetic sensors @ $5
Rogowski coil @ $150
UM250k ultramic @ $250
Multi-channel analog-to-digital converter $200
3D printer @ $650