Image Steganalysis Scheme Based on Transfer Learning Using CNN, BiLSTM

Steganography, the practice of covert information exchange, allows data to be hidden within various embedding mediums such as image, audio, video, and text files. This paper focuses specifically on the use of images for steganography and steganalysis. We propose a novel hybrid model aimed at enhancing the detection accuracy of context-aware steganography techniques. The primary objective of this research is to address the limitations of current methods by integrating advanced preprocessing and deep learning techniques. Our proposed model leverages transfer learning to reduce training time and resource consumption, employing Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks as recognizers. The model operates in three distinct phases: (1) preprocessing with Fast Fourier Transform (FFT), Hilbert transform, and Discrete Wavelet Transform (DWT); (2) feature extraction using CNNs for spatial characteristics and BiLSTM for contextual information; and (3) classification by aggregating outputs with a feed-forward neural network (FFNN). The proposed model is evaluated on the BOSSbase and CIFAR-256 image sets using the HILL, S-Uniward, and WOW context-aware steganography algorithms. Experimental results demonstrate that our technique outperforms existing methods in terms of detection accuracy and computational efficiency. The model achieves accuracies of 95.9%, 95.1%, and 91.1% for WOW, S-Uniward, and HILL on the BOSSbase dataset, respectively. On the CIFAR-256 dataset, accuracies of 93.9%, 95.35%, and 90.07% are attained for WOW, S-Uniward, and HILL, respectively.