B-INN: an invertible neural network framework for credential-bound image ownership verification

The rapid expansion of image datasets for artificial intelligence has increased ownership disputes caused by unauthorized collection, redistribution, and reuse. Conventional watermarking and steganographic methods can embed identifying information into images, but successful extraction alone may not be sufficient for ownership verification when the extraction model or embedded evidence is obtained by an unauthorized party. This article proposes B-INN, a credential-bound image ownership verification framework that combines Invertible Neural Network-based data hiding with blockchain-managed Verifiable Credential records. B-INN embeds a credential hash as image-bound ownership evidence and verifies the recovered evidence against the registered credential record. The experimental results show that B-INN recovers the decoded credential hash without error under clean conditions, achieving a bit accuracy of 1.00, a bit error rate of 0, and an exact recovery rate of 1.00. Under moderate degradation conditions, including JPEG compression, blur, and noise, small recovery errors are observed in some cases. To handle such imperfect recovery, B-INN applies a threshold-based ownership verification procedure. Under the selected threshold and the defined threat model, false ownership claim evaluation shows zero false acceptance across wrong credential, random hash, replay, different-owner, and model exposure scenarios. These results indicate that B-INN does not rely on extraction alone as ownership authority, but links recoverable image-bound evidence with registered credential records for post-distribution image ownership verification.