eduzhai > Applied Sciences > Engineering >

Dynamically Mitigating Data Discrepancy with Balanced Focal Loss for Replay Attack Detection

  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: It becomes urgent to design effective anti-spoofing algorithms for vulnerableautomatic speaker verification systems due to the advancement of high-qualityplayback devices. Current studies mainly treat anti-spoofing as a binaryclassification problem between bonafide and spoofed utterances, while lack ofindistinguishable samples makes it difficult to train a robust spoofingdetector. In this paper, we argue that for anti-spoofing, it needs moreattention for indistinguishable samples over easily-classified ones in themodeling process, to make correct discrimination a top priority. Therefore, tomitigate the data discrepancy between training and inference, we propose toleverage a balanced focal loss function as the training objective todynamically scale the loss based on the traits of the sample itself. Besides,in the experiments, we select three kinds of features that contain bothmagnitude-based and phase-based information to form complementary andinformative features. Experimental results on the ASVspoof2019 datasetdemonstrate the superiority of the proposed methods by comparison between oursystems and top-performing ones. Systems trained with the balanced focal lossperform significantly better than conventional cross-entropy loss. Withcomplementary features, our fusion system with only three kinds of featuresoutperforms other systems containing five or more complex single models by22.5 for min-tDCF and 7 for EER, achieving a min-tDCF and an EER of 0.0124and 0.55 respectively. Furthermore, we present and discuss the evaluationresults on real replay data apart from the simulated ASVspoof2019 data,indicating that research for anti-spoofing still has a long way to go.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×