Committed Computation Traces: Hybrid Verification of Neural Network Inference Between Probabilistic Replay and Zero-Knowledge Proofs
Verifying outsourced neural network inference has two established options and a wide gap between them. Re-running a sample of the work on independent hardware is cheap but catches a dishonest job only as often as it samples one. Proving every arithmetic step cryptographically is near-certain but, on production-scale models, costs a large multiple of the inference itself.
This paper proposes Committed Computation Trace Verification (CCTV), which sits between them. The executor commits to its internal computation trace through hash-based commitments before it learns what will be challenged; the verifier then opens and re-executes a random subset of inter-layer constraints, selected with a server-held secret so the choice cannot be predicted. Because falsifying one layer invalidates every constraint downstream of it, sampling a few constraints catches far more than the sampling rate alone would suggest.
The expected detection probability is derived in closed form as 1 - (L-k)/((k+1)L), which for a 24-layer transformer with 5 sampled constraints gives 86.8 percent. An implementation on BAAI/bge-large-en-v1.5 (335M parameters, 24 layers) measured 86.7 percent, with 100 percent detection of cached replay and model substitution and no false rejections on honest executions. The paper states an explicit adversary taxonomy, including the case the construction does not improve against: an adversary who recomputes every downstream layer honestly leaves one violated constraint, and is caught at roughly the plain sampling rate.