RFD 013 idea
Network Exhaustion Protocol
The Idea
Minimize unexplained network traffic leaving a cluster so undeclared workloads (e.g., secret training on an inference cluster) become detectable. Combine with workload declaration, inline network tap, random/full logging, and random recomputation.
For inference: single network cable to cluster, all traffic passes through tap, declared workloads are randomly recomputed to verify logs match computation.
Why It Matters
Network bandwidth is a scarce resource that’s hard to hide. If all bandwidth is accounted for by declared workloads, undeclared workloads have nowhere to send their outputs. This is a chokepoint-based verification strategy.
Open Questions
- How much bandwidth overhead does verification add?
- Can attackers compress or steganographically hide outputs?
- How to handle legitimate but variable network usage?