Calibrated steganalysis of mp3stego in multi-encoder scenario

摘要

Experimenting with different mp3 encoders shows subtle differences in their outputs indicating they have been implemented differently. If this characteristic is not addressed properly, it could degrade performance of mp3 steganalysis. Additionally, calibration is a powerful technique which has not found its true potential in mp3 steganalysis. This paper tries to fill these gaps. First, we propose a new set of calibrated features based on quantization step. To that end, we show that quantization steps of mp3 bit stream and steganography-induced noise are respectively band-limited and wide-band signals. We use this observation and apply a low pass filter on quantization steps of mp3 bit stream and then use the result for calibrating of features. Then, we present our analysis on different encoders and show that their statistical properties are quite dissimilar for all fields of mp3 bitstream. We use this observation and propose an encoder classifier that have near-perfect accuracy. This accuracy is achieved after testing a wide range of feature selection methods and finally using genetic algorithm due to its superior performance. Our simulations show that the proposed method achieved accuracy of 95.8% and 95.2% at 3.12% of maximum embedding rate in the single and multi-encoder scenarios, respectively. (C) 2018 Elsevier Inc. All rights reserved.