Research On Key Technologies Of Deep Learning-Based Audio Steganalysis

Digital steganography is an essential means of ensuring communication security,but it also provides convenience for malicious communication.As a countermeasure to steganography,steganalysis aims to detect whether secret information is embedded in the signal,which has important research significance in this era of increasing attention to information security.Audio is one of the commonly used digital media in people's daily life.The popularity of audio applications and software has promoted the development of audio steganography and steganalysis technologies.Nowadays,deep learning-based steganalysis has achieved excellent performance under laboratory conditions.However,there is a large amount of diversity and complexity audio data in complex big data forensics scenarios.The steganography algorithms and cover models used by attackers are usually unknown,which can impact the performance of the steganalysis distinctly.In this paper,we focus on digital audio,and research on some of the critical problems of audio steganalysis based on deep learning.The main contributions of the paper are summarized as follows:Firstly,we propose a convolutional neural network structure suitable for end-to-end audio steganalysis.We design a high-order difference-based high pass filter at the beginning of the network.In this way,an automatic learning method of audio steganographic noise feature expression is effectively established.In the case of low embedding rate,the model parameter fine-tuning strategy is introduced to improve the detection performance.The experimental results show that the detection performance of the proposed model can surpass the existing deep learning-based audio steganalysis method under a variety of steganography algorithms and embedding rates.Compared to the traditional handcrafted steganalysis paradigm,our proposed method also has a certain performance improvement.Secondly,we propose a domain adaptation-based audio steganalysis when facing the cover source mismatch problem in complex big data forensics scenarios.By introducing the domain discriminator in the traditional supervised steganalysis paradigm,and making it adversarial training with the feature extractor,which aims to extract the effective steganalysis domain-invariant and steganalytic feature.Experimental results on heterogeneous audio steganography datasets generated by different recording equipment show that the proposed method can effectively alleviate the detection performance degradation of the classifier caused by cover source mismatch.Finally,we propose a universal audio steganalysis framework when facing the steganography algorithm mismatch in big data forensics scenarios.By introducing multi-task learning and the anomaly detection mechanism,the generalized performance of the general steganalysis method in mismatched scenarios can be effectively improved.The experiment proves that this work can not only improve the detection effect of dense audio in the mismatch of the steganography algorithm but also distinguish the specific type of steganography algorithm.Besides,our proposed method can also effectively distinguish the algorithms not included in the training data.