The number of channels is one of the important criteria in regard to digital audio quality. Generally, stereo audio with two channels can provide better perceptual quality than mono audio. To seek illegal commercial benefit, one might convert a mono audio system to stereo with fake quality. Identifying stereo-faking audio is a lesser-investigated audio forensic issue. In this paper, a stereo faking corpus is first presented, which is created using the Haas effect technique. Two identification algorithms for fake stereo audio are proposed. One is based on Mel-frequency cepstral coefficient features and support vector machines. The other is based on a specially designed five-layer convolutional neural network. The experimental results on two datasets with five different cut-off frequencies show that the proposed algorithm can effectively detect stereo-faking audio and has good robustness.

Although neural network-based speech recognition models have enjoyed significant success in many acoustic systems, they are susceptible to be attacked by the adversarial examples. In this work, we make first step towards using generative adversarial network (GAN) for constructing the targeted speech adversarial examples. Specifically, we integrate the target speech recognition network with GAN framework, which can then be formulated as a three-party game. The generator in GAN aims at generating perturbation that could make the target network misclassified to a specific target, while simultaneously fooling the discriminator treating the adversarial example as a beguine one. The discriminator is to distinguish the crafted examples from the geniue samples. The classification error of the target network is back-propagated via gradient flow to the generator for updating. The target network is responsible for back-propagating the classification error via gradients to the generator for updating, but the target network itself is freezed. With the carefully designed network architecture, loss function and training strategy, we successfully train a generator that could generate the adversarial perturbation for a given speech clip and a target label. Experiential results show that the generated adversarial examples could effectively fool the state-of-the-art speech classification networks, while attaining an acceptable auditory perception quality. In addition, our proposed method runs much faster than the prevalent optimization-based schemes. To facilitate reproducible research, codes, models and data are publicly available at https://github.com/winterwindwang/SpeechAdvGan.

Audio steganography aims to exploit the human auditory redundancy to embed the secret message into cover audio, without raising suspicion when hearing it. However, recent studies have shown that the existing audio steganography can be easily exposed with the deep learning based steganalyzers by extracting high-dimensional features of stego audio for classification. The existing GAN-based steganography approaches mainly studied in images cover, less work is conducted on audio cover. In addition, though a few GAN-based audio steganography methods have been proposed, they still have room for improvements in perceptual quality and undetectability. In this work, we propose an audio steganography framework that could automatically learn to generate superior steganographic cover audio for message embedding. Specifically, the training framework of the proposed framework consists of three components, namely, generator, discriminator and trained deep learning based steganalyzer. Then the traditional message embedding algorithm LSBM, is employed to embed the secret message into the steganographic cover audio to obtain stego audio, which is delivered to the trained steganalyzer for misclassifying as cover audio. Once the adversarial training is completed among these three parties, one can obtain a well-trained generator, which could generate steganographic cover audio for subsequent message embedding. In the practice of our proposed method, the stego audio is produced by embedding the secret message into the steganographic cover audio using a traditional steganography method. Experimental results demonstrate that our proposed audio steganography can yield steganographic cover audio that preserves a quite high perception quality for message embedding. We have compared the detection accuracies with the existing audio steganography schemes as presented in our experiment, the proposed method exhibits lower detection accuracies against the state-of-the-art deep learning based steganalyzers, under various embedding rates. Codes are publicly available at https://github.com/Chenlang2018/Audio-Steganography-using-GAN.

With the widespread availability of cell-phone recording devices, source cell-phone identification has become a hot topic in multimedia forensics. At present, the research on the source cell-phone identification in clean conditions has achieved good results, but that in noisy environments is not ideal. This paper proposes a novel source cell-phone identification system suitable for both clean and noisy environments using spectral distribution features of constant Q transform (CQT) domain and multi-scene training method. Based on the analysis, it is found that the identification difficulty lies in different models of cell-phones of the same brand, and their tiny differences are mainly in the middle and low frequency bands. Therefore, this paper extracts spectral distribution features from the CQT domain, which has a higher frequency resolution in the mid-low frequency. To evaluate the effectiveness of the proposed feature, four classification techniques of Support Vector Machine (SVM), Random Forest (RF), Convolutional Neural Network (CNN) and Recurrent Neuron Network-Long Short-Term Memory Neural Network (RNN-BLSTM) are used to identify the source recording device. Experimental results show that the features proposed in this paper have superior performance. Compared with Mel frequency cepstral coefficient (MFCC) and linear frequency cepstral coefficient (LFCC), it enhances the accuracy of cell-phones within the same brand, whether the speech to be tested comprises clean speech files or noisy speech files. In addition, the CNN classification effect is outstanding. In terms of models, the model is established by the multi-scene training method, which improves the distinguishing ability of the model in the noisy environment than single-scenario training method. The average accuracy rate in CNN for clean speech files on the CKC speech database (CKC-SD) and TIMIT Recaptured Database (TIMIT-RD) databases increased from 95.47% and 97.89% to 97.08% and 99.29%, respectively. For noisy speech files with seen noisy types and unseen noisy types, the performance was greatly improved, and most of the recognition rates exceeded 90%. Therefore, the source identification system in this paper is robust to noise.