iSeal was proposed in November 2025 by a research team from Stevens Institute of Technology and the University of Houston. The related research findings were published in the paper "iSeal: Encrypted Fingerprinting for Reliable LLM Ownership Verification"It has been accepted by AAAI 2026".

iSeal is the first method to reliably verify fingerprint recognition when a model thief controls a suspected LLM end-to-end. First, it introduces an external encoder to decouple the fingerprint from the model, preventing reverse engineering via weighted access alone. Second, its cryptographic design features strong diffusion and obfuscation, ensuring that even instantaneous response pairs do not leak useful information, thus preventing forgetting and fingerprint inference under collusion. To prevent response control, the researchers employ similarity-based verification rather than fragile exact matching. Finally, the error correction module provides provable robustness, allowing recovery even if the preceding module fails.