Steganography with Statistical Models of Image Noise Residual

Steganography alters innocuously looking cover objects in order to communicate in secrecy. This manuscript focuses on steganography in digital images, arguably the most popular and most studied cover objects. The current focus of steganography is on content-adaptive schemes that are realized through minimizing a distortion function designed to focus the attention of the embedding on highly textured regions of images that are hard to model and where the embedding is less detectable. The actual embedding is done through efficient coding schemes. As interesting as this whole paradigm of embedding by minimizing distortion might seem, distortion in not detectability. It is only linked heuristically through the design of the distortion function.;One of the contributions of this dissertation is to formulate this problem through statistical hypothesis testing theory by modeling image noise residuals as a sequence of independent and quantized zero-mean Gaussian random variables. Within this model, the most secure steganographic approach is the one that Minimizes the Power of the Most Powerful Detector (MiPOD) built to distinguish between cover and stego objects. To the best of the author's knowledge, the proposed model-based embedding scheme, MiPOD, is the first embedding scheme of this kind which has a comparable security with respect to current state of the art in content-adaptive steganography. This dissertation also looks into many interesting implications of having a model-based approach for steganography and steganalysis. The model-based detector is used to assess the performance of current feature-based steganalysis schemes and their optimality. A new detectability-limited sender is proposed that adjusts the embedded payload inside each image up to a certain prescribed level of detectability. Furthermore, for the first time, the proposed detector enables us to measure the secure payload size for a single image for a certain prescribed detectability.;Recently, it has been shown that the detection power of feature-based steganalysis can be improved by reducing the redundancy in the extracted feature vectors by focusing the attention of the feature extractor more towards the heavily embedded regions inside each image, hence selection-channel-aware feature sets. This dissertation, among other contributions, presents a systematic approach to study the effect of having inaccuracies between steganographer's activities and steganalyst's presumed assumptions about those activities, e.g., the embedding payload, and having access to the cover source. It is proposed to model these inaccuracies as four different types of Warden with different levels of knowledge about the selection channel to assess the security of state-of-the-art embedding schemes under these different settings.;Finally, this dissertation uses the proposed model-based schemes to reformulate the problem of batch steganography and pooled steganalysis. The most powerful detector, aware of the spreading strategies used by Alice inside each communication bag of images, is built as a matched filter and further simplified using a practical estimation approach. Furthermore, three intuitive payload spreading strategies are proposed with roots inside both model-based and content-adaptive steganography.