ZK-AgriTwin: A Blockchain-Enabled Framework with Verifiable Edge Intelligence and Adaptive Committee BFT for Autonomous Agri-Food Traceability

Abstract

Although the supply chain for agri-food has been digitalised and allows the visibility of the provenance of the information, the traditional blockchain-based systems are limited to ensuring the integrity of the information after it has been submitted to the blockchain ledger, but they are not able to establish that the information was reliable at the time of its initial physical recording. This can therefore be permanently recorded even if the blockchain validation is successful, such as manipulated sensor measurements, incorrect quality assessment, compromised edge devices, inconsistent custody declarations, privacy leakage, and inefficient Byzantine consensus. ZK-AgriTwin proposes a privacy-preserving cyber-physical traceability system which combines multimodal digital twins, verifiable edge intelligence, cross-modal consistency checking, zero-knowledge verification of inference, encrypted content-addressed storage and adaptive committee Byzantine fault tolerance. Every agricultural batch is captured as a digital twin that is continually developed based on product pictures, physicochemical data, environmental data, storage data, digital provenance, and custody data. Multimodal intelligence running at the lightweight end of the data source checks quality conditions and decides if the different observations are consistent with a physical condition before accepting the traceability event. Inference evidence is kept secret using zero-knowledge-verifiable computation, allowing validators to verify approved inference decisions without knowing the raw images of the products, confidential measurements from the sensors or even the attributes of the products which are commercially sensitive. Consensus is done by dynamically formed Byzantine-resilient committees with members drawn based on cryptographically verifiable randomness and reliability, availability, communication quality and observed adversarial risk. Large Multimodal Records are stored in the encrypted content-addressed storage, while state commitments, proof references, anomaly decisions, custody events and consensus certificates are anchored on-chain. Evaluation is organized around open-access multimodal litchi observations, measurements collected from commercial cold storage, and independent fruit imagery, evaluating predictive reliability, cross-modal anomaly detection, digital-twin synchronization, proof overhead, consensus scalability, Byzantine resilience, communication efficiency and storage reduction. The resulting framework moves blockchain traceability beyond simply recording data submitted to the blockchain to verifiable, physical-to-digital provenance, assuming evidence credibility, computational integrity, confidentiality and distributed agreement prior to irreversible blockchain ledger commitment.