A Paradigm Shift for Computational Excellence from Traditional Machine Learning to Modern Deep Learning-Based Image Steganalysis
Neelam Swarnkar, Arpana Rawal, Gulab Patel · 2021
Steganography is a method of hiding data in digital files in order to conceal its very existence, whereas steganalysis is the practice of detecting the presence of the secret (hidden) messages in the multimedia files (text, image, audio, and video files). The success of the steganalytic model highly depends upon its ability to extract effective feature representations present in the data set. This chapter spans across two comparative paradigms of feature extractions in the history of research timeline in steganalysis realms. The traditional paradigm, which is usually a two-step machine learning (ML) approach comprising feature extraction followed by a classification step. This approach relies on good quality; however, handcrafted features of high dimensionality based on enormously relevant statistical computations over image characteristics eventually leads to improved classification accuracies with respect to conventional steganalytic systems. The modern paradigm is a recent breakthrough in efficient feature representation and automated feature learning approach since 2015, coined as Deep Learning (DL) approaches. some noteworthy models to mention: Deep-Boltzmann Machines, Deep-autoencoders, and Convolution Neural Networks (CNNs) were introduced, leading to improved classification accuracies as against traditional ML-based steganalytic systems. These DL models unify the two steps, feature extraction and classification, under a single architecture, along with optimizing the dimensionality of extracted feature sets. This chapter discusses weight initialization methods for avoiding vanishing and exploding gradient problems, activation functions for introducing nonlinearity, pooling for reducing the dimensionality of features, optimizers for increasing training speed of the model, residual learning (ResL) for strengthening weak stego signals, batch normalization for faster convergence of the network, spatial pyramid pooling for handling arbitrary size images, and data augmentation and transfer learning for better performance of the steganalyzers.