Toward Hidden Data Detection via Local Features Optimization in Spatial Domain Images

Technology advancements made machine learning algorithms crucial to solving complex problems. Deep learning, a machine learning paradigm to design convolutional neural networks (CNNs), achieves promising performance in detecting confidential data, known as steganalysis. However, the existing steganalysis CNNs have not achieved optimal performance detecting accuracy and network stability. In this research, we propose a new approach within CNN to improve the secret data detection accuracy by optimizing the local features in the feature extraction stage of the spatial domain images. The performance is evaluated using the Break Our Steganographic System Base (BOSSBase) dataset with two standard adaptive steganography algorithms employing low payload capacities of 0.2 and 0.4 bits per pixel. The experimental results outperform the results of the previously published works in accuracy and network stability. The detection accuracy is improved in a range between 2.1% to 3.6%.