An Image Compression Based Technique to Watermark a Neural Network
While neural networks have made considerable progress in the area of digital representation, training of neural models requires an enormous data and time. It is well known that the use of trained models as initial weights often leads in less training error than un-pre-trained neural networks. We propose in this paper a digital watermarking system for neural networks. We formulate a new challenge: the integration of watermarks into neural networks through discrete cosine transform (DCT) based approach. For discrete wavelet transform (DWT)-based digital image watermarking algorithms, additional performance enhancements could be obtained by combining DWT with DCT. Throughout the neural networks, we also describe specifications, embedded conditions, and attack forms of watermarking. The technique presented here does not affect the network performance in which a watermark is positioned as the watermark is embedded while the host network is being trained. Finally, we perform detailed image data experiments to demonstrate the potential of neural networks watermarking as the basis for this research attempt.