Study On Robust Videos Watermarking And Deep Neural Networks Watermarking

With the rapid development of communication technology and the popularization of mobile smart devices,the sharing of short videos has become a popular way of socializing.For copyrights owners of videos,it is an urgent problem to ensure that their videos will not be edited into short videos for arbitrary distribution without authorization.The redundancy that video data can provide makes it possible to provide space for the embedding of digital watermarks even after being encoded by a compression standard.Therefore,the digital watermarking technology,which can not only ensure that the visual quality of the video is not significantly degraded,but also protect the copyright of the video,can effectively solve this problem.In addition,due to the rapid development of deep learning and widely spreading of its applications,deep neural network(DNN),which requires extensive resources for training,is also recognized as an intellectual property that needs copyright protection.Digital watermarking can also be extended to the field of DNN.Without affecting the performance of the DNN model,the copyright protection of the DNN model can be implemented by embedding a digital watermark in it.This dissertation mainly studies the robust video watermarking and DNN watermarking algorithms,the performances of these watermarking algorithms are evaluated by fidelity,robustness and capacity.The main contributions are as follows:1.A robust H.264/AVC video watermarking algorithm without intra distortion drift is proposed.Although the existing anti-distortion drift scheme guarantees the visual quality of the video,it decreases the watermark capacity and lacks of consideration on the robustness of the watermark.By analyzing the intra prediction mode of H.264/AVC,the 4 × 4 luminance block is re-classified to improve the capacity.To further improve the fidelity of the watermarking algorithm,the mean square error before and after the watermarking is leveraged to select the embedding position of the watermark under each classification of the luminance block.The preprocessing of the watermark by combining the spread spectrum technology and the error correction code effectively improves the robustness of the watermark.Eventually,the watermarking is completed by modifying the quantized discrete cosine transform(DCT)coefficients that meet the conditions of the selected luminance block.Experimental results for this algorithm show that the embedded video has good subjective and objective visual quality even when the watermark capacity is increased,and it has good robustness in resisting attacks such as re-encoding and low-pass filtering attack.2.A new DNN watermarking algorithm that leverages on Spread Transform Dither Modulation(ST-DM)is proposed.To ensure the capacity of the watermark,the proposed white-box multi-bit watermarking algorithm selects the weight full of redundancy as the embedding position.By analyzing ST-DM algorithm,it is found that it can be combined with weight to achieve watermarking.Spread transformation is realized by projecting weights onto multiple directions of the customized projection matrix,which provide a higher payload with a lower impact on network accuracy while retaining a satisfactory level of robustness.The conventional dither modulation function is a nonlinear function,which cannot satisfy the back-propagation mechanism of DNN.Therefore,this paper proposes a dither modulation function suitable for DNN,which can quantify the projection value to the [0,1] interval.The regularization term composed of ST-DM is added to the loss function of the DNN model,and the disordered quantized value sequence is approximated to the watermark sequence to be embedded through model training.Experimental results show that the algorithm has good performance in terms of fidelity,robustness and watermark capacity.3.A feature-map-based dynamic DNN watermarking algorithm is proposed.The watermark embedded in the weight must be embedded and extracted as a whole,while the dynamic watermark based on the feature map can output different watermarks in correspondence to a set of predefined inputs,which is more flexible.Thanks to the high redundancy of feature maps,the new approach can achieve a very high payload.To improve the robustness of the watermark,the defined spreading matrix makes the watermark evenly distributed in the entire feature map.The watermarking is achieved by training quantized production between feature map and the spreading matrix.In addition,this proposed method analyzes the influence on performances of this algorithm by utilizing different types of trigger inputs.Users can choose the type of the input that triggers the watermark according to their own needs.The experimental results show that,compared with existing schemes,although the robustness and fidelity of the algorithm are slightly reduced,the watermark capacity has been greatly improved,which can reach more than four times that of existing schemes.