zkRNN: Zero-knowledge proofs for Recurrent Neural Network Inference

Zero-knowledge proofs (ZKPs) enable verification of computations without revealing private information and have recently been applied to neural-network inference. Existing approaches, however, primarily target feedforward and convolutional architectures and do not naturally extend to recurrent neural networks (RNNs), whose sequential dependencies and shared parameters introduce new verification challenges. We present zkRNN, a zero-knowledge framework for verifiable RNN inference that allows a prover to demonstrate the correctness of model outputs without revealing model parameters, intermediate hidden states, or private inputs. Our approach extends the GKR protocol with a recursive sum-check framework tailored to recurrent computation. We introduce an arithmetic-circuit representation that captures hidden-state transitions and weight sharing across time steps and employ recursive proof composition to aggregate proofs generated at each recurrent step. As a result, verifier time and proof size remain polylogarithmic in the size of a single-step circuit and independent of sequence length. We further support streaming inference through incremental proof generation and aggregation, enabling low-latency verification for sequential workloads. Experimental results demonstrate efficient proof generation and sequence-length-independent verification across realistic RNN configurations. For sequence lengths up to T = 256, zkRNN maintains a proof size of roughly 18 KB, while prover time scales linearly with sequence length and verification completes in under 10 seconds. These results broaden the applicability of verifiable machine learning to sequential models and time-dependent data.