Anti-forensics of fake stereo audio using generative adversarial network

Fake-quality audio detection is an important branch in the field of digital audio forensics. Resampling and recompression are the two typical operations to achieve fake audio quality, in which an audio with low sampling/bit rate can be converted to one with higher sampling/bit rate pretending to be in high quality. Stereo-faking is another fake-quality operation, with which a mono audio can be converted into a stereo one. To detect the stereo-faking, a few forensic methods have been proposed. Little consideration, however, has been given to the security of these methods themselves. To expose the weakness of these stereo-faking detectors, an anti-forensic framework based on generative adversarial network is proposed. The fake stereo audio is created by generating a new channel audio based on a mono audio. Skip connection is adopted to ensure the quality of the generated audio. Considering that stereo application scenarios are mostly music and film recording, a large number of music and film recordings are downloaded from the Internet as our datasets. Use these datasets to train our model. The anti-forensic samples generated by the model are used to attack the most effective fake stereo audio detectors. Experimental results show that the generated fake stereo audio of music can significantly reduce its detection accuracy from about 99-30%, and the false acceptance rate can increase from 0.08% to about 69%. The fake stereo audio generated from the film recording can significantly reduce its detection accuracy from about 99-1.7%, and the false acceptance rate can increase from 0.02% to about 98%.

Automated summary

Fake-quality audio detection, specifically in the context of stereo-faked audio, is an important field in digital audio forensics. This study proposes an anti-forensic framework based on generative adversarial network to expose the weaknesses of stereo-faking detectors. By generating fake stereo audio using a mono audio, the researchers demonstrate that detection accuracy can significantly decrease while the false acceptance rate increases.