PWNN Paper: Neural Networks as Remote Fault Injectors to Steal AES Keys From Cloud FPGAs
chaumian · x · 2026-09-18
Researchers at Karlsruhe Institute of Technology present PWNN (Power-Wasting Neural Network) in IACR TCHES: a technique that weaponizes the inherent switching behavior of neural network operations under adversarial inputs to induce voltage fluctuations and timing faults on shared cloud FPGAs, enabling Differential Fault Analysis and successful AES key recovery. The paper systematically explores architectures and input patterns, and warns that existing verification and structural DRC are blind to such attacks embedded in benign-looking circuits, as cloud FPGA demand for LLM workloads surges.
More from Safety
- Ben Todd: optimists have no rigorous model proving advanced AI agents are safe either — ben_j_todd · 2026-09-18
- UK government commits £115M to AI biosecurity and agentic AI incident response — ben_j_todd · 2026-09-18
- Pedro Domingos: Any proposal to slow down AI is unworkable and only helps those who game it — pmddomingos · 2026-09-18
- PCP v0.1: A Draft Protocol for Purpose-Bound, Revocable AI Agent Authority — sierracatalina · 2026-09-18
- Sentdex slams HF hack evals as lazy, questions airgapped-comms dismissal — cephaloform · 2026-09-18
- Brussels probes OpenAI agents accused of hijacking a German wiki after 18,000 messages — emmanuelvivier · 2026-09-18