Study On Double Compression Detection Of AMR And AAC Audio

With the advent of the era of mobile Internet and 4G mobile communication, as well as the extensive application of smart mobile devices, more and more people are using smart mobile devices to record audios with compressed format. The tamper detection of compressed audio has drawn more and more attention from researchers. However, the current studies are mainly focused on the MP3 format. In this work, we investigate the double compression detection of two other compressed audio formats, namely the AMR(Adaptive Multi-Rate) and the AAC(Advanced Audio Coding) formats.The main contributions are summarized as follows:1. For AMR audio, a double compression detection method based on deep learning is put forward. First, the AMR file is decompressed to WAV file from which a feature vector is extracted. The feature vector is composed of the average low-frequency subband energy ratio, the average subband energy ratio, the bispectrum features and the long-term LPC(Linear Predictive Coding) coefficients. At last, the SAE(Stacked Auto Encoder) is used as the core network to judge whether the audio has experienced double compression operation. To further improve the performance of the algorithm, a voting strategy is employed. Specifically, the WAV file is split into several clips, the detecting results of every clip are used to vote for the detecting result of the whole file. Experimental results show that the algorithm can detect AMR double compression operation effectively. Moreover, for those AMR files transcoded from low bitrate to high bitrate, the original bitrate can also be detected with our algorithm.2. According to the frequency domain masking effect of AAC encoding, a method based on SAE is proposed to detect the double compression history of AAC audio. The proposed method can not only judge whether an AAC audio has experienced the double compression operation, but also detect the original bitrate of the AAC audio that has been transcoded from low bitrate to high bitrate. The AAC audio is first decompressed to WAV format, the spectrum of which is split into several subbands. A feature vector that consists of the normalized energies of each subband is fed to an SAE network for classification. A voting strategy is used to improve the performance of the algorithm. In the end, a method based on spectrum observation is proposed to determine whether an AAC audio has been transcodied from low bitrate to high bitrate and the original bitrate as well.