| Information hiding technology like steganography has been witnessing a monumental growth in a wide range of practical applications.Unfortunately,it could be used by terrorist organizations to conduct criminal activities,so there is an imperative desire to detect malicious steganography,which is known as steganalysis.As deep learning has greatly promoted the development of computer vision technology,researchers and enthusiasts alike have been evaluating the effectiveness of neural networks in the field of steganalysis to discover the malicious use of steganography.In this study,a multi-scale feature fusion method and the attention mechanism were designed to construct a neural network for audio steganalysis.(1)A multi-scale parallel multi-branch network was used to sample the characteristic signals with different granularities to adjust the receptive field.The attention mechanism was introduced to perform weight learning on feature information after multi-scale processing,and a weighted combination of multiple feature representations was then obtained to improve the performance of the CNN model.The experimental results show that the proposed neural network integrating with a multi-scale feature fusion method and the attention mechanism could detect malicious steganography with 97.35% accuracy.Experimental results show that the proposed research scheme confirms the effectiveness of multiscale feature fusion methods and attention mechanisms in the field of VoIP audio steganalysis.(2)In order to further improve the training efficiency and detection efficiency,this paper also proposes a Steganalysis Transformer(SAT)VoIP speech steganography analysis method based on Transformer neural network.This method first encodes the relative position of the features extracted from the VoIP voice signal,and designs the mapping layer in combination with a multi-scale method;then uses the improved Transformer to perform feature extraction to realize steganalysis;proposes that the SAT method can obtain global features from the shallow layer,and learn High-quality intermediate features.Experiments show that the SAT method proposed in this paper has superior performance,and the steganalysis accuracy rate of VoIP speech is 96.45%.The training efficiency and detection efficiency are about three times higher than the previous method. |
