Dark Silicon: Unmasking GPU Threats in the Age of AI
summary
GPUs are silently abused for crypto‑mining and model theft: VRAM‑resident malware (eg. Jellyfish) and side‑channel snooping leave no OS logs, while a distillation campaign ran ~24,000 fake accounts and ~16 million queries to exfiltrate model behavior. Live demos show resource hijacks that raise GPU utilization from ~63% to ~94% without EDR alerts and recovery only after manual inspection. Detection and forensics use existing tools—nvidia-smi, centralized NVML telemetry, and a kernel‑level inspector CUP—plus audit, patch, and isolation steps.
AI-written summary · about
chapters (10)
- 0:00GPU threats introduction
- 1:05Hidden GPU cryptojacking heist
- 3:55Why GPU attacks are missed
- 7:06AI model theft via distillation
- 12:47STRIDE threat model for GPUs
- 16:30Live GPU attack demos
- 18:35Tools to monitor GPUs
- 21:47Audit patch isolate strategy
- 23:31Final takeaways
- 24:55Q&A
related talks
29mBarraCUDA: GPUs Do Leak DNN Weights
20mTrusted Hardware for AI Workloads: Extending Confidential...
43mThe GPU Utilization Problem: What’s Going Wrong and How to Solve It
13mDemocratizing AI with Open Source Compute
55mMore Tales from the Dark Side: How AI is the Bad Guys New Friend
18mTelekine: Secure Computing with Cloud GPUs
25mAgentic Zero Trust: at Rest, in Transit, and at Runtime
26mWrap-Up Session Panel Discussion: Quick Takes