2
most popular audio format, MP3 audio can be easily manipulated by some audio
editing software, such as CoolEdit, GoldWave and so on. Thus the double MP3
compression detection for audio forensics becomes more significant than before.
While image and video double compression detection have attracted many
scholars’ attention, research on double MP3 compression detection is rare, espe-
cially for the detection of the same bitrate. Yang et al. [2, 3] considered that the
number of the MDCT coefficients between -1 and 1 in double compressed MP3
audio is less than that in single compressed MP3 audio. The method can detect the
double compressed MP3 audio transcoded from lower bitrate to high bitrate, but
the results show the low accuracy rate on the other opposite situation. Liu and Qi-
ao et al. [4, 5] observed the statistical characteristics of the MDCT coefficients
which exceed a certain threshold, a support vector machine is applied for classifi-
cation, experiment results demonstrate that the algorithm can correctly detect
whether the double compressed MP3 audio is transcoded from high bitrate to low-
er bitrate or from low bitrate to higher bitrate.
In summary, although many researchers have been devoted to the double MP3
compression under different bitrates, there is still a blank in the double com-
pressed audio detection with the high bitrate, which means both the first and the
second compressed bitrate are higher than that of normal. In this case, the encoder
has enough bits to store the audio information, and the loss of the audio infor-
mation is tiny after the first and the second compression, which result in the slight
change of the effective features between them [6]. All the traditional algorithms of
the detection are based on the statistical features of quantized MDCT coefficients.
This paper presents a novel algorithm which utilizes the statistical probability of
the Huffman table index as the characteristics to detect double MP3 compression.
The rest of this paper is organized as follows. In section 2, the procedures of
the Huffman coding in MP3 encoding are reviewed briefly. The statistical features
of the Huffman code table index are observed between single and double com-
pressed MP3 audios in section 3. Experiments of detecting double MP3 compres-
sion are implemented based on the extracted features in section 4. Finally, the
conclusions and future work are summarized in section 5.
2. Application of Huffman code table
As a lossy encoder in audio compression, MP3 encoder can maximize the filtered
information that human auditory system is unable to perceive, without affecting
the audio quality of the original hearing. Not only that, the encoder make use of
the lossless Huffman coding, which compress the audio volume as tiny as possible
so as to facilitate the network transmission.
MP3 encoding is in units of frames. Each frame contains two granules, and
each granule contains two channels in stereo music. Each channel includes 576
quantized MDCT coefficients for Huffman coding at the end of the MP3 encoding