Fingeprinting GPU workloads with eBPF
summary
Uses eBPF kprobes on the NVIDIA kernel driver to record timestamps for 41 kprobe hooks across 60 seconds per run, tested on A10 and V100 GPUs against TensorFlow training, LLaMA and NVIDIA inference, and a coin-miner. Converts per-function time series into DTW distances using dynamic time warping and reduces those 41 distances to a single norm, producing stable fingerprints per GPU and workload; legitimate runs cluster within ~5–40 distance while the miner sits ~400 away, demonstrating a working PoC with hardware-sensitive fingerprints and a possible patent filing.
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 1:48GPU workload detection challenge
- 5:15Proof of concept setup
- 8:47Fingerprinting with dynamic time warping
- 13:27Future directions and Q&A
related talks
51meBPF Warfare - Detecting Kernel & eBPF Rootkits with Tracee
30mxprobes: Hybrid User/Kernel eBPF Probes for Cross-Layer Observability
33meBPF: exploring use case of BPF kernel infrastructure
28mDark Silicon: Unmasking GPU Threats in the Age of AI
31meBPF and Wasm: Unifying Userspace Extensions With Bpftime
19mDTrace and eBPF: new challenges
26mChecking your cryptography usage with eBPF
40mExploring Runtime Security and Forensic using eBPF