Adaptive spatial image steganography and steganalysis using perceptual modelling and machine learning

Image steganography is a method for communicating secret messages under the cover images. A sender will embed the secret messages into the cover images according to an algorithm, and then the resulting image will be sent to the receiver. The receiver can extract the secret messages with the predefined algorithm. To counter this kind of technique, image steganalysis is proposed to detect the presence of secret messages. After many years of development, current image steganography uses the adaptive algorithm for embedding the secrets, which automatically finds the complex area in the cover source to avoid being noticed. Meanwhile, image steganalysis has also been advanced to universal steganalysis, which does not require the knowledge of the steganographic algorithm. With the development of the computational hardware, i.e., Graphical Processing Units (GPUs), some computational expensive techniques are now available, i.e., Convolutional Neural Networks (CNNs), which bring a large improvement in the detection tasks in image steganalysis. To defend against the attacks, new techniques are also being developed to improve the security of image steganography, these include designing more scientific cost functions, the key in adaptive steganography, and generating stego images from the knowledge of the CNNs. Several contributions are made for both image steganography and steganalysis in this thesis. Firstly, inspired by the Ranking Priority Profile (RPP), a new cost function for adaptive image steganography is proposed, which uses the two-dimensional Singular Spectrum Analysis (2D-SSA) and Weighted Median Filter (WMF) in the design. The RPP mainly includes three rules, i.e., the Complexity-First rule, the Clustering rule and the Spreading rule, to design a cost function. The 2D-SSA is employed in selecting the key components and clustering the embedding positions, which follows the Complexity-First rule and the Clustering rule. Also, the Spreading rule is followed to smooth the resulting image produced by 2D-SSA with WMF. The proposed algorithm has improved performance over four benchmarking approaches against non-shared selection channel attacks. It also provides comparable performance in selection-channel-aware scenarios, where the best results are observed when the relative payload is 0.3 bpp or larger. The approach is much faster than other model-based methods. Secondly, for image steganalysis, to tackle more complex datasets that are close to the real scenarios and to push image steganalysis further to real-life applications, an Enhanced Residual Network with self-attention ability, i.e., ERANet, is proposed. By employing a more mathematically sophisticated way to extract more effective features in the images and the global self-Attention technique, the ERANet can further capture the stego signal in the deeper layers, hence it is suitable for the more complex situations in the new datasets. The proposed Enhanced Low-Level Feature Representation Module can be easily mounted on other CNNs in selecting the most representative features. Although it comes with a slightly extra computational cost, comprehensive experiments on the BOSSbase and ALASKA#2 datasets have demonstrated the effectiveness of the proposed methodology. Lastly, for image steganography, with the knowledge from the CNNs, a novel postcost-optimization algorithm is proposed. Without modifying the original stego image and the original cost function of the steganography, and no need for training a Generative Adversarial Network (GAN), the proposed method mainly uses the gradient maps from a well-trained CNN to represent the cost, where the original cost map of the steganography is adopted to indicate the embedding positions. This method will smooth the gradient maps before adjusting the cost, which solves the boundary problem of the CNNs having multiple subnets. Extensive experiments have been carried out to validate the effectiveness of the proposed method, which provides state-of-the-art performance. In addition, compared to existing work, the proposed method is effcient in computing time as well. In short, this thesis has made three major contributions to image steganography and steganalysis by using perceptual modelling and machine learning. A novel cost function and a post-cost-optimization function have been proposed for adaptive spatial image steganography, which helps protect the secret messages. For image steganalysis, a new CNN architecture has also been proposed, which utilizes multiple techniques for providing state of-the-art performance. Future directions are also discussed for indicating potential research.Image steganography is a method for communicating secret messages under the cover images. A sender will embed the secret messages into the cover images according to an algorithm, and then the resulting image will be sent to the receiver. The receiver can extract the secret messages with the predefined algorithm. To counter this kind of technique, image steganalysis is proposed to detect the presence of secret messages. After many years of development, current image steganography uses the adaptive algorithm for embedding the secrets, which automatically finds the complex area in the cover source to avoid being noticed. Meanwhile, image steganalysis has also been advanced to universal steganalysis, which does not require the knowledge of the steganographic algorithm. With the development of the computational hardware, i.e., Graphical Processing Units (GPUs), some computational expensive techniques are now available, i.e., Convolutional Neural Networks (CNNs), which bring a large improvement in the detection tasks in image steganalysis. To defend against the attacks, new techniques are also being developed to improve the security of image steganography, these include designing more scientific cost functions, the key in adaptive steganography, and generating stego images from the knowledge of the CNNs. Several contributions are made for both image steganography and steganalysis in this thesis. Firstly, inspired by the Ranking Priority Profile (RPP), a new cost function for adaptive image steganography is proposed, which uses the two-dimensional Singular Spectrum Analysis (2D-SSA) and Weighted Median Filter (WMF) in the design. The RPP mainly includes three rules, i.e., the Complexity-First rule, the Clustering rule and the Spreading rule, to design a cost function. The 2D-SSA is employed in selecting the key components and clustering the embedding positions, which follows the Complexity-First rule and the Clustering rule. Also, the Spreading rule is followed to smooth the resulting image produced by 2D-SSA with WMF. The proposed algorithm has improved performance over four benchmarking approaches against non-shared selection channel attacks. It also provides comparable performance in selection-channel-aware scenarios, where the best results are observed when the relative payload is 0.3 bpp or larger. The approach is much faster than other model-based methods. Secondly, for image steganalysis, to tackle more complex datasets that are close to the real scenarios and to push image steganalysis further to real-life applications, an Enhanced Residual Network with self-attention ability, i.e., ERANet, is proposed. By employing a more mathematically sophisticated way to extract more effective features in the images and the global self-Attention technique, the ERANet can further capture the stego signal in the deeper layers, hence it is suitable for the more complex situations in the new datasets. The proposed Enhanced Low-Level Feature Representation Module can be easily mounted on other CNNs in selecting the most representative features. Although it comes with a slightly extra computational cost, comprehensive experiments on the BOSSbase and ALASKA#2 datasets have demonstrated the effectiveness of the proposed methodology. Lastly, for image steganography, with the knowledge from the CNNs, a novel postcost-optimization algorithm is proposed. Without modifying the original stego image and the original cost function of the steganography, and no need for training a Generative Adversarial Network (GAN), the proposed method mainly uses the gradient maps from a well-trained CNN to represent the cost, where the original cost map of the steganography is adopted to indicate the embedding positions. This method will smooth the gradient maps before adjusting the cost, which solves the boundary problem of the CNNs having multiple subnets. Extensive experiments have been carried out to validate the effectiveness of the proposed method, which provides state-of-the-art performance. In addition, compared to existing work, the proposed method is effcient in computing time as well. In short, this thesis has made three major contributions to image steganography and steganalysis by using perceptual modelling and machine learning. A novel cost function and a post-cost-optimization function have been proposed for adaptive spatial image steganography, which helps protect the secret messages. For image steganalysis, a new CNN architecture has also been proposed, which utilizes multiple techniques for providing state of-the-art performance. Future directions are also discussed for indicating potential research