Perceptual Hashing of Deep Convolutional Neural Networks for Model Copy Detection

In recent years, many model intellectual property (IP) proof methods for IP protection have been proposed, such as watermarking and fingerprinting. However, with the increasing number of models transmitted deployed on Internet, quickly finding suspect among thousands model-sharing platforms GitHub is in great demand, which concurrently triggers new security problem copy detection protection. As an important part system, task has not received enough attention. Due to high computational complexity, both fingerprinting lack capability efficiently find suspected infringing tens millions models. this article, inspired by hash-based image retrieval methods, we introduce a novel mechanism: perceptual hashing convolutional neural networks (CNNs). The proposed algorithm can convert weights CNNs fixed-length binary hash codes so that lightly modified version similar code original model. By comparing similarity pair between query test library, versions be retrieved efficiently. To best our knowledge, first deep network Specifically, select based compression theory, then calculate normal statistics (NTS) segments weights, finally encode NTS features into codes. experiment performed library containing 3,565 indicates scheme superior performance.