基于时域统计特征的音频内容取证新算法

A Novel Audio Content Forensics Scheme Based on Time Domain Statistical Characteristic

针对现有音频内容取证算法采用二值图像作为辨识水印所带来的安全隐患,以及基于音频内容或特征生成的辨识水印稳定性不高,易被常规信号处理操作淹没的问题,提出了一种新的基于时域统计特征的音频内容取证算法。通过对音频信号时域统计平均值进行非均匀量化生成辨识水印。理论和实验结果表明通过该方法生成的辨识水印能够抵抗常规信号处理操作,稳定性高。生成的辨识水印存储于认证中心,组建辨识水印库。对音频内容进行取证时,将由该音频生成的辨识水印与从水印库中提取的对应辨识水印进行比对,即可对待取证音频的真实性、完整性进行鉴定。该取证方法操作简便,对不同类型音频均能实现篡改定位,对常规音频信号处理操作的鲁棒性高,有效扩大了基于内容音频取证算法的应用范围。

Many previous audio content forensics schemes adopt binary image as identifying watermark, which introduces security holes to forensics systems. On the other hand, partial content-based or feature-based identifying watermarks have feeblish stability and may be damaged under various signal processing operations. To overcome these problems, a novel audio content forensics scheme based on time domain sta-tistical characteristic is proposed in this paper. The statistical average value of continuous audio samples is used to generate identifying watermark by non-uniform quantization. Theoretical analysis and experimental results show that the generated identifying watermark is robust against various signal processing operations. Various identifying watermarks generated from different audio signals are stored at CA （ Center of Authenti-cation）. When authenticating the veracity and integrity of audio content, firstly identifying watemark is generated from the to be detected audio, then corresponding identifying watermark is extracted from data-base of CA, finally the two identifying watermarks for audio content forensics are compared. The proposed forensics scheme has lower computation complexity, and the ability of tamper localization and tolerance a-gainst common signal processing operations are excellent. It greatly expands the applicability of content- based audio forensics scheme.