Parallel Methods For Stego Image Feature Extraction

Steganography detection techonology can play a role to prevent malicious communication by detecting whether carriers contain hidden information, thus it becomes an effective countermeasure against illegal behavior that abuses steganography techniques. Steganography blind detection is a pattern recognition process and it comprises two key steps:carrier feature extraction and classifier construction. Feature extraction is to extract some sensitive and discernable statistics from carriers as feature vectors, which is the key step in steganography blind detection. Facing mass image media appearing on the Internet, how to improve the time efficiency of image steganography blind detection is a problem to be urgently deal with. Stego image feature extraction has large amount of calculation, so it is the most time consuming part in detection process. Meanwhile, it gradually becomes a trend that constructing steganography feature vector with high dimension, which leads to the currently used serial method encounter insurmountable time performance bottlenecks. In this thesis, using multicore CPU and general purpose GPU as parallel processors, we focus on the parallelization implementation of stego image feature extraction in image steganography detection. The main work of this thesis includes:1. A parallel method for stego image feature extracton on multicore CPU:On multicore CPU platform, by taking advantage of the little correlation in stego image feature extraction calculation process, we propose a feature extraction method that uses thread-level task parallelism, which firstly constructs a lock-free task queue for task threads, secondly reduces thread synchronization overhead and finally solves false sharing issue, optimize memory conflict and sets thread affinity scheduling to improve performance. In this way, we achieve the goal that using multicore CPU computing resource accelerates stego image feature extraction. The experiments indicate that our method is better to utilize multicore CPU performance than the original OpenMP method.2. A parallel method for SRM feature extraction algorithm on multicore CPU: Image feature extract algorithm base on rich model can extract over ten thousand dimensional feature vector, and this calculation process is very time comsuming. So, we use SRM feature extraction algorithm, represented the feature extract algorithm base on rich model, to explain the parallelization of the rich model feature extraction algorithm. On one hand, we use SIMD instruction level parallel method to reduce the serial time of the SRM algorithm; on the other hand, we combine the instruction level parallel method with the thread level parallel method introduced in "Parallel method for stego image feature extracton on multicore CPU" to achieve further parallel optimization on multicore CPU system. Results of the experiments show the new method improves the computational efficiency of SRM algorithm significantly.3. A parallel method base on GPU for SRM feature extraction algorithm:as the image size increases, the computing time of SRM feature extraction algorithm rapidly growths in non-linear way. Therefore, we implement the SRM algorithm with OpenCL on AMD GPU. By means of the detail analysis of GPU architecture characteristic and the meticulous optimization of SRM algorithm, the GPU parallel method achieve better calculation time on large size images than the parallel method proposed in "A parallel method for SRM feature extraction algorithm on multicore CPU".Finally, a comprehensive analysis is made among the three parallel methods above and we explain the applicable situation of each method.