Verifying declared workloads from network-tap logs

Develop analysis methods that, given sampled inter-accelerator communication logs (including GPU kernels) captured by mutually vetted network taps in an AI compute cluster, can verify whether the cluster executed only the declared high-level code and no undeclared workloads.

Background

The report proposes off-chip network taps that log sampled inter-accelerator data, including kernels, to check for discrepancies between actual communications and declared workloads. This would support Subgoals 1.A and 2.A in the verification framework by detecting false declarations or undeclared uses on declared clusters.

However, turning these logs into a reliable determination of whether the cluster executed only the declared high-level code requires new analysis techniques that are robust to adversaries and sampling limitations.