Improved steganalysis technique based on least significant bit using artificial neural network for MP3 files

• Main Author: Alarood, Ala Abdulsalam Solyiman
• Format: Thesis
• Language: English
• Published: 2017
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.utm.my/id/eprint/81789/1/AlaAbdulsalamSolyimanPFC2017.pdf; http://eprints.utm.my/id/eprint/81789/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:126087

MP3 files are one of the most widely used digital audio formats that provide a high compression ratio with reliable quality. Their widespread use has resulted in MP3 audio files becoming excellent covers to carry hidden information in audio steganography on the Internet. Emerging interest in uncovering such hidden information has opened up a field of research called steganalysis that looked at the detection of hidden messages in a specific media. Unfortunately, the detection accuracy in steganalysis is affected by bit rates, sampling rate of the data type, compression rates, file track size and standard, as well as benchmark dataset of the MP3 files. This thesis thus proposed an effective technique to steganalysis of MP3 audio files by deriving a combination of features from MP3 file properties. Several trials were run in selecting relevant features of MP3 files like the total harmony distortion, power spectrum density, and peak signal-to-noise ratio (PSNR) for investigating the correlation between different channels of MP3 signals. The least significant bit (LSB) technique was used in the detection of embedded secret files in stego-objects. This involved reading the stego-objects for statistical evaluation for possible points of secret messages and classifying these points into either high or low tendencies for containing secret messages. Feed Forward Neural Network with 3 layers and traingdx function with an activation function for each layer were also used. The network vector contains information about all features, and is used to create a network for the given learning process. Finally, an evaluation process involving the ANN test that compared the results with previous techniques, was performed. A 97.92% accuracy rate was recorded when detecting MP3 files under 96 kbps compression. These experimental results showed that the proposed approach was effective in detecting embedded information in MP3 files. It demonstrated significant improvement in detection accuracy at low embedding rates compared with previous work.