Deepfake Detection System for the ADD Challenge Track 3.2 Based on Score Fusion

Authors: Yuxiang Zhang, Jingze Lu, Xingming Wang, Zhuo Li, Runqiu Xiao, Wenchao Wang, Ming Li, Pengyuan Zhang

Published: 2022-10-13 08:04:29+00:00

Comment: Accepted by ACM Multimedia 2022 Workshop: First International Workshop on Deepfake Detection for Audio Multimedia

AI Summary

This paper describes a deepfake audio detection system submitted to the Audio Deep Synthesis Detection (ADD) Challenge Track 3.2, focusing on score fusion. The system leverages score-level fusion of multiple Light Convolutional Neural Network (LCNN) based models with various input features and online data augmentation. The authors analyze the reasons for the limited performance improvement of score fusion, attributing it to model overfitting and low correlation of scores on out-of-distribution test data.

Abstract

This paper describes the deepfake audio detection system submitted to the Audio Deep Synthesis Detection (ADD) Challenge Track 3.2 and gives an analysis of score fusion. The proposed system is a score-level fusion of several light convolutional neural network (LCNN) based models. Various front-ends are used as input features, including low-frequency short-time Fourier transform and Constant Q transform. Due to the complex noise and rich synthesis algorithms, it is difficult to obtain the desired performance using the training set directly. Online data augmentation methods effectively improve the robustness of fake audio detection systems. In particular, the reasons for the poor improvement of score fusion are explored through visualization of the score distributions and comparison with score distribution on another dataset. The overfitting of the model to the training set leads to extreme values of the scores and low correlation of the score distributions, which makes score fusion difficult. Fusion with partially fake audio detection system improves system performance further. The submission on track 3.2 obtained the weighted equal error rate (WEER) of 11.04\\%, which is one of the best performing systems in the challenge.

Key findings

The proposed LCNN-based score fusion system achieved a competitive Weighted Equal Error Rate (WEER) of 11.04% in the ADD Challenge Track 3.2. Online data augmentation methods, particularly noise and reverberation addition, significantly improved system robustness. However, score fusion showed limited improvement due to model overfitting, leading to highly confident but poorly correlated scores on out-of-distribution test data.

Approach

The problem is solved using an ensemble approach based on score-level fusion of several LCNN models. These models are trained with diverse front-end features like low-frequency STFT and Constant Q transform, and incorporate online data augmentation methods. The system also includes a partially fake audio detection component for enhanced robustness.

Datasets

ADD 2022 Challenge datasets (training, development, track 1 adaptation set, track 3.2 test set), MUSAN, RIRs, ASVspoof 2019 LA (for score fusion analysis).

Model(s)

Light Convolutional Neural Network (LCNN), Bidirectional Long Short-Term Memory (BLSTM), AM-Softmax Loss, Center Loss.

Author countries

China

← Previous