The verifiable AI technology ‘vCNN’, jointly developed by the research team of Professor Kim Ji Hye from the Department of Electronic Engineering at our university and the research team of Professor Oh Hyun Ok from the Department of Information Systems at Hanyang University, has won the 2024 Best Paper Award from the international journal IEEE Transactions on Dependable and Secure Computing (TDSC). This research received high praise for its ability to prove the reliability of AI inference results without disclosing input data or the model itself.

As AI expands into medical diagnosis, financial fraud detection, and public services, the need to ensure the accuracy of result calculations, protect sensitive data (medical images, transaction records), and keep model weights—a core corporate asset—confidential has grown. However, existing methods either exposed personal information or trade secrets for accuracy verification, or faced excessive computational time when applying cryptographic techniques, hindering large-scale service adoption.

The joint research team's developed vCNN (Verifiable Convolutional Neural Network) resolves these core bottlenecks. vCNN utilizes zero-knowledge proofs (zk-SNARKs) to provide a short, concise proof that ‘the AI calculated accurately according to the prescribed procedure’ without disclosing the input or model. Notably, it redesigned the proof mechanism for convolution—CNN's core operation—reducing computational complexity from O(l·n) (where l is kernel size and n is data size) to O(l·n). This achieved approximately 20-fold faster proof generation for the MNIST model and 18,000-fold faster for the VGG16 model, while mathematically proving its security.