Audio steganalysis using multi-scale feature fusion-based attention neural network

Abstract

Deep learning techniques have shown promise in audio steganalysis, which involves detecting the presence of hidden information (steganography) in audio files. However, deep learning models are prone to overfitting, particularly when there is limited data or when the model architecture is too complex relative to the available data for VoIP steganography. To address these issues, new deep-learning approaches need to be explored. In this study, a new convolutional neural network for audio steganalysis, incorporating a multi-scale feature fusion method and an attention mechanism, was devised to enhance the detection of steganographic content in audio signals encoded with G729a. To improve the network's adaptability, a multi-scale parallel multi-branch architecture was employed, allowing characteristic signals to be sampled with varying granularities and adjusting the receptive field effectively. The attention mechanism enables weight learning on the feature information after multi-scale processing, capturing the most relevant information for steganalysis. By combining multiple feature representations using a weighted combination, the deep learning model's performance was significantly enhanced. Through rigorous experimentation, an impressive accuracy rate of 94.55% was achieved in detecting malicious steganography. This outcome demonstrates the efficacy of the proposed neural network, leveraging both the multi-scale feature fusion method and the attention mechanism.