A protection method of trained CNN model with a secret key from unauthorized access

In this paper, we propose a novel method for protecting convolutional neural network models with secret key set so that unauthorized users without the correct cannot access trained models. The enables us to protect not only from copyright infringement but also functionality of model any noticeable overhead. We introduce three block-wise transformations generate learnable transformed images: pixel shuffling, negative/positive transformation, and format-preserving Feistel-based encryption. Protected are by using images. results experiments CIFAR ImageNet datasets show performance protected was close non-protected when correct, while accuracy severely dropped an incorrect given. demonstrated be robust against various attacks. Compared state-of-the-art protection passports, proposed does have additional layers in network, therefore, there is no overhead during training inference processes.