Skip to content
Atikul Islam Munna logo Home
Back to projects

Research · 2026

ChainAttest: Cross-Chain Verification of Privacy-Preserving ML Attestations from Permissioned to Public Blockchains

Manuscript in preparation

Abstract

Blockchain bridges are built to move assets, but machine-learning provenance needs a different payload: typed evidence about a model’s identity, its dataset and training commitments, and how it performed in evaluation. ChainAttest is a protocol and reference implementation that carries these attestations across chains.

A committee authenticates each source record after a finality window, while ML-specific Groth16 proofs verify its meaning on the destination chain, so source authenticity and semantic verification stay separate. Evaluation claims are privacy-preserving: the destination chain sees a proved verdict over blinded commitments, never the score or the sample counts. A coordinator builds, signs, submits and recovers the whole flow end to end, for both EVM and permissioned Fabric-style sources.

Contributions

  • Typed relay packages for model attestations and evaluation claims, verified by destination contracts with replay protection and revocation handling.
  • A blinded evaluation circuit that proves a minimum sample size and a pass threshold over Poseidon commitments, without revealing the underlying counts.
  • Committee authentication for both EVM and permissioned Fabric-style sources, kept independent of the semantic proofs.
  • An operational coordinator with SQLite persistence, restart recovery and an authenticated HTTP signer boundary.

Evaluation

  • Gas compared against a trusted-bridge baseline and a naive digest anchor, on both the EVM and the Fabric source paths.
  • Scaling sweeps over Merkle tree depth and evaluation batch count, plus a repeated 30-run end-to-end benchmark.
  • A run on the Sepolia public testnet matched local devnet gas within 24 gas (under 0.005%).

Research areas

  • Blockchain
  • Zero-Knowledge Proofs
  • AI Provenance
  • Privacy-Preserving Evaluation
  • Smart Contracts