Forensic and anti-forensic analysis of indoor/outdoor classifiers based on acoustic clues

This paper addresses the problem of identifying the class of the environment where an audio recording was taken. We focus on distinguishing between indoor and outdoor speech recordings, and we propose a set of classifiers that provide a support for the forensic analyst in verifying the authenticity of audio content. The classifiers rely on acoustic clues extracted from the reverberant signal, namely the reverberation time (RT₆₀) and MFCC/LMSC feature vectors. We conducted several experiments, aimed at analyzing the algorithms from both the forensic and anti-forensic perspective. To do so, we devised a methodology for manipulating the signals in order to pretend that outdoor contents were recorded indoor, and vice-versa. Experimental results confirm the effectiveness of the proposed methods, which achieve high classification ac curacy. The anti-forensics analysis reveals that attacks have moderate success rates, and severely depend from the classi fication algorithm adopted by the analyst.