Parallel Implementation Of Steganalysis On A Graphic Processing Unit

With the steganography unceasing changes and billions of multimedia files continuously appear in the Internet, two coping measures should be taken by steganalysis. First, propose a universal steganalysis algorithm with good generalization ability, in order to coping with unknown steganography algorithm. Second, improve steganalysis algorithm's real-time characteristic and powering performance, in response to hundreds of millions of multimedia files. In this thesis, a parallel steganalysis idea based on GPU is proposed, it uses image as the research object, and from the view of real-time and powering, GPU with the parallel programming ability is used as the parallel machine, which is launched by NVIDIA.The parallel implementation of image steganalysis based on GPU gains a very good runing speed.The main work is accomplished in thesis as following:1. According to the characteristic "large data, weak correlation" of image, we present a parallel method based on pixel level. At the same time, we put forward the parallel evaluation indicators and evaluation method of GPU as the parallel machine.2. Two batch images scheduling methods base on CUDA framework are proposed. Method 1, combine batch images into one image and bind the image to a texture. We could avoid more program branch using clamping addressing mode, which improves the parallel computing speed greatly. Method 2. combine batch images into one image and transfer the image into the GPU global memory, which reduces the communication frequency greatly. More over, this method adopts two kinds of images combination modes, more flexible than the former.3. The parallel RS steganalysis algorithm for BMP image based on GPU is realized. By means of mining the parallelism for RS algorithm in pixel level, point and neighborhood parallel and feature extraction parallel, experiments of a single image/batch images based on one device and batch images based on multi-device for RS parallel steganalysis are designed. Accoding to the time-consumption statistics of experiments above, we evaluate the parallel performance of parallel RS steganalysis algorithm.4. The parallel NJD steganalysis algorithm for JPEG image based on GPU is realized. We design different parallel algorithms arroding to the parallelism of NJD feature extraction and SVM classification. The method of time-consumption statistics and parallel evaluation is same to 3.