Efficient privacy-preserving scheme for secure neural network inference

• Proposes a three-stage privacy-preserving inference scheme combining homomorphic encryption and secure multi-party computation to protect both user data and model parameters. • Introduces network parameter merging to reduce multiplication levels and ciphertext–plaintext operations, enhancing efficiency. • Develops a fast convolution algorithm that significantly boosts computational performance in ciphertext inference. • Achieves at least 11% reduction in online stage linear operation time compared to state-of-the-art methods, cutting inference time and communication overhead.