Effects of incremental training on watermarked neural networks

Abstract

Deep learning has achieved extraordinary results in many different areas, ranging from autonomous driving [1], medical devices [2] to speech recognition and natural language processing [3]. Generating a high-performance neural network is costly in aspects of time, computational resources, and expertise, making the models valuable intellectual property (IP). As a result, there has been a notable growth in attention and investments in the paradigm of machine learning. In recent years, watermarking methods have been developed in order to protect the Intellectual Property Rights (IPR) of neural networks, and many schemes have successfully prevented adversaries from stealing such models. However, little has been studied on how Incremental Training would affect the persistence of watermarks in such watermarking schemes. This investigation aims to discover the effects of Incremental Training on in existing watermarking schemes. Keywords: Intellectual Property Rights (IPR), Watermarking, Incremental Training