CREDIT: Ownership verification of deep neural networks against model extraction attacks

The researchers propose CREDIT, a method for ownership verification of deep neural networks against model extraction attacks, achieving current best performance on mainstream datasets across multiple domains and tasks. The method uses mutual information to quantify the similarity between DNN models, proposes a practical verification threshold, and based on this threshold provides rigorous theoretical guarantees for ownership verification. The authors note that although many defense strategies exist, ownership verification of suspicious models still lacks rigorous theoretical guarantees, and this work addresses exactly this gap. The implementation code is publicly available on GitHub (https://github.com/LabRAI/CREDIT）。