Novel Blind Detection of LSB Steganography Based on Kernel FDA

With rapid progress of watermarking steganalysis becomes urgent to prevent transmission of dangerous information. A novel method to detect the existence of LSB steganography blindly is proposed. Image segmentation is employed firstly to utilize local properties. According to the segmentation method used, the formation of one component is also the formation of its minimum spanning tree (MST). By graph theory the difference of MST's node degree of corresponding components in clean and stego images is presented. Considering that information is usually hidden in complex textures and obscure region image complexity is defined on many parameters related to the essence of image. Kernel fisher discriminant analysis (FDA) is employed to decide whether LSB steganography information exists or not. Simulation results show that the proposed method performs much better than RS method at different embed-ing rate and false positive rate.