LSBR Speech Steganalysis Based on Percent of Equal Adjacent Samples

Abstract

Wide utilization of audio files has attracted the attention of cyber-criminals to employ this media as a cover for their concealed communications. As a countermeasure and to protect cyberspace, several techniques have been introduced for steganalysis of various audio formats, such as MP3, VoIP, etc. The combination of machine learning and signal processing techniques has helped steganalyzers to obtain higher accuracies. However, as the statistical characteristics of a normal audio file differ from the speech ones, the current methods cannot discriminate clean and stego speech instances efficiently. Another problem is the high numbers of extracted features and analysis dimensions that drastically increase the implementation cost. To tackle these, this paper proposes the Percent of Equal Adjacent Samples (PEAS) feature for single-dimension least-significant-bit replacement (LSBR) speech steganalysis. The model first classifies the samples into speech and silence groups according to a threshold which has been determined through extensive experiments. It then uses an MLP classifier to detect stego instances and determine the embedding ratio. PEAS steganalysis detects 99.8% of stego instances in the lowest analyzed embedding ratio — 12.5% — and its sensitivity increases to 100% for the ratios of 37.5% and above.