Telekine: Secure Computing with Cloud GPUs
summary
Telekine demonstrates that GPU kernel timing leaks enough information to classify ImageNet images by observing kernel durations, then builds a protect-and-remote design that moves control-plane code to the client and uses API remoting plus a new data-oblivious streams primitive to mask timing. Prototype runs MXNet training with a 10–22% slowdown (Galois graph workloads 18–41%), requires a constant ~533 Mbps dummy traffic rate, and prevents the demonstrated timing classifier from recovering labels.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:16Trusting the cloud provider is difficult
- 1:30Avoiding trust in the cloud provider is difficult
- 2:25Introduce TEEs to isolate computation
- 3:56TEEs have limitations
- 8:43Information gained from kernel execution time
- 11:14Data-oblivious streams
- 13:29We compare Telekine to an insecure baseline: running on the GPU server without protections
- 14:50MXNet neural net training (Real WAN)
- 16:33Telekine: Secure Computing with Cloud GPUs
related talks
28mDark Silicon: Unmasking GPU Threats in the Age of AI
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud
29mCloud and Edge: Price, Performance and Privacy Considerations in IOT
20mConfidential Computing within an AI Accelerator
24mStrictly Confidential: Use Cases & Demos of AI Workflows in TEEs
25mAgentic Zero Trust: at Rest, in Transit, and at Runtime
19mSOTER: Guarding Black-box Inference for General Neural Networks at the Edge
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads