A CNN Based Audio Steganalysis Algorithm by Manual Feature Extraction and Result Merging

With the rapid development of Internet technology, IP-based voice transmission technology has emerged. While bringing convenience to people, it also brings many security risks. The criminals using VoIP voice transmission protocol in compressed domains to transmit secret information has brought great challenges to social security. In this paper, for the pitch steganography algorithm and the quantized index modulation audio steganography algorithm of complementary neighbor vertex based on G.729A encoding, an audio steganalysis algorithm based on manual feature extraction and convolutional neural network is proposed. By combining manually extracted features with convolutional neural networks, it is possible to achieve effective detection of both the quantized index modulation audio steganography algorithm of complementary neighbor vertex and the pitch-based steganography algorithm in the VoIP compressed domain. Specifically, the algorithm proposed in this paper firstly extracts manual features from the G.729A speech segment (including two manual features extracted by the pitch steganography algorithm and three manual features extracted by the quantized index modulation audio steganography algorithm with complementary neighbor vertex). After using audio steganography algorithm to steganography audio samples, the five extracted manual features have been changed to vary degrees. Therefore, these five manual features can be used as one of the basis for judging whether the audio samples contain secret information. Then, after extracting the five manual features, this paper designs two different convolutional neural networks for the pitch steganography algorithm and the quantized index modulation audio steganography algorithm with complementary neighbor vertex. The two extracted manual features for the pitch steganography algorithm and the three manual features for the quantized index modulation audio steganography algorithm based on complementary neighbor vertex are input into the two different convolutional neural networks, respectively. Immediately afterwards, the two convolutional neural networks will further extract and discriminate the input manual features, and obtain the steganalysis results based on the pitch audio steganography algorithm and the quantized index modulation audio steganography algorithm with complementary neighbor vertex, respectively. Finally, according to a designed fusion rule, the network merges the two discriminant results to obtain the final discriminant result, that is, the network discriminates whether the input audio sample contains steganographic information. In summary, the algorithm proposed in this paper extracts features manually from the audio samples encoded by G.729A, and combines the manually extracted features with the convolutional neural network, which can effectively perform steganalysis and detection on the pitch audio steganography algorithm and the quantized index modulation audio steganography algorithm with complementary neighbor vertex in the VoIP compression domain. The experimental results show that in detecting both the pitch steganography algorithm and the quantized index modulation audio steganography algorithm with complementary neighbor vertex at the same time, the detection accuracy rate of the proposed audio steganalysis algorithm based on manual feature extraction and the convolutional neural network proposed in this paper can reach 86.2% (when the embedding rate is 100% and the audio sample duration is 0.1s). Compared with the existing excellent steganalysis algorithms, the algorithm proposed in this paper has achieved state-of-the-art detection results when the audio duration is shorter. © 2021, Science Press. All right reserved.