CNN based Image Steganalyzer

• Steganography is the process of hiding data into public digital medium (carrier) for secret communication. Image steganography has become a widely used phenomena as the digital images can hide the secret data without much distortion, due to the presence of redundant information in the images. The image in which the secret data is hidden is termed as stego image. • The objective of Steganography is to hide the very presence of hidden information, rather than content of information. • Steganalysis is the method by which the presence of hidden data is detected in an image or carrier. • The objective of Steganalysis is detection of presence of secret message, rather than the contents of it. • Every steganography method employs a special mechanism to embed the secret data in the images. Therefore, it leaves a distinct pattern on the stego images. • By identifying these traces of distortion, which are caused by embedding of the data, Steganalysis aims to detect whether a given image contains a secret message or not. • Steganalysis methods are broadly classified into two categories. • Targeted Steganalysis. • Universal (Blind) Steganalysis 1. Targeted steganalysis is a method of detection devised for a particular steganographic method. That means the steganalyst has the access to the method of steganography by which data is hidden. (it is also called Signature Steganalysis) 2. In Universal steganalysis method the steganalyst does not have the knowledge of embedding steganographic method. Hence it is also called Blind steganalysis. Perhaps it is less accurate than targeted steganalysis but it is very effective against new methods of steganography. Image steganalysis is a very challenging task because the stego signal introduced to images is rather weak. Traditional steganalysis consisted of two parts. 1. Computing/extracting features from images 2. Training a classifier, such as SVM or Ensembled based on these features to distinguish between cover and stego objects. The big question is How to extract powerful features to capture the traces caused by steganographic operations? Examples for handcrafted features are Image quality metrics, SPAM, RichModel(SRM) etc., key limitations . Firstly, though the classification module is trainable, the feature extraction module is manually designed, which requires a great deal of human intervention and expertise. To obtain a good feature representation, the steganalyzer needs to carefully consider as many useful statistics as possible. However, it is difficult because typical image statistics are greatly influenced by factors like in-camera processing, image preprocessing and the variation of image contents. In addition, with the rapid development of advanced steganography, the statistical changes caused by embedding operations are much harder to model, which makes the handcrafted feature extraction even more difficult. Secondly, the feature extraction and classification steps are separated in traditional steganalysis systems. Once some useful information is lost in the feature extraction, it can not be recovered in the classification step. Recently, Steganalysis tasks are being done using CNNs The major advantages of CNNs is that two process conventional steganalysis is done in a single step. Conventional steganalysis where the features are handpicked, requires a great domain knowledge on image statistics, where as CNNS automatically extract the features by training. This feature makes the CNNs an obvious choice for steganalysis. Though CNNs have been very successful with image classification and recognition, in case of steganalysis, the scenario changes because the noise caused by the steganography in the cover image is very weak and difficult to identify. We have to design a CNN based steganalyzer for image classification into stego or cover which is novel in its idea and competes with current state of the art techniques.