一种改进的音频内容认证算法

An Improved Content Based on Audio Authentication Algorithm

介绍了一种基于内容的数字音频认证算法,主要目的在于突破现有音频认证算法仅能判断音频是否受到篡改、而难以定位一个乃至多个具体受篡改位置的局限性,提供更具意义的认证结果。现有的算法通过鲁棒性哈希和半脆弱水印来进行内容的认证,一般采用固定时间分帧并且缺乏重新同步的机制,因此当多处剪切、插入这类失同步的攻击发生时,这些算法都无法在攻击的结束位置重新同步从而导致认证结果有较高的漏警率。文章所述的算法通过提取音频中的关键锚点,以此将音频划分为一系列不等长片段,利用动态时间规整的方法完成同步,最后由机器学习训练得到的模型协助音频的内容认证。实验结果证明了该方法在多于一次的失同步攻击的情况下仍然保持良好的认证效果。

In this paper an improved content-based audio authentication algorithm is proposed. Most of the existing works either use robust hash methods or semi-fragile watermarks to verify the perceptual integrity in fixed time frames, and thus can hardly tell the manipulation position if attacks occur in multiple places. Our approach focuses on providing the justified content authentication results even one or more de-synchronizing attacks （cropping or inserting） take place. In particular, the audio piece is segmented into a series of non-overlapped sound events to be aligned with the reference audio, and then according to the authentication result from an supervised model, attack positions can be detected. Experiment shows encouraging content authentication results compared to the pervious approaches using fixed framing and without re-synchronizing methods.