Agree on the Model, Verify the Inference: GKR Protocols for HND-Based Transformer Inference

As Transformer models grow in scale, outsourcing inference to cloud providers becomes necessary for resource-constrained clients. However, this creates a trust gap: clients cannot easily verify if the provider actually used the registered model or completed the full computation without re-executing the model themselves, which negates the performance benefits of outsourcing. This paper asks: can a client verify the execution of a registered Transformer backbone without storing dense weights or performing expensive matrix re-computation?

The authors introduce GKR-HND, a protocol that leverages the Homomorphic–Nonhomomorphic Decomposition (HND) of Transformers. HND separates the model into a polynomial backbone and compact non-polynomial correction modules. By focusing on the polynomial backbone, the authors apply the GKR (Goldwasser-Kalai-Rothblum) protocol, which allows a verifier to check the output of a layered arithmetic circuit with significantly less work than full execution. The system further optimizes this by separating cryptographic verification from public terminal computation, delegating the latter to an assigned worker whose response is cryptographically bound to the request.

The study demonstrates that GKR-HND effectively verifies the polynomial backbone of HND models. By using a block-local GKR proof, the verifier reduces the computational burden, avoiding the need for dense-matrix replay. The protocol ensures that if the prover or worker deviates from the registered computation, the verifier will detect the discrepancy through the proof transcript or the signed worker response. Experimental results confirm that the proof path and delegated public computations are valid and computationally efficient for pretrained HND models.

This work provides a practical path toward verifiable AI services. By offloading the verification burden from the client and avoiding full re-execution, GKR-HND makes it feasible to audit cloud-based model inference. This is a critical step for ensuring model integrity in sensitive applications where users must be certain that the specific, registered model was used to generate their results.