AdfNet:一种基于多样化特征的自适应深度伪造检测网络

AdfNet:An Adaptive Deep Forgery Detection Network Based on Diverse Features

视频篡改造成的危害一直在危及人们的生活,这使深度伪造检测技术逐渐得到广泛关注和发展。然而,目前的检测方法由于使用了不灵活的约束条件,无法有效捕获噪声残差;此外,也忽略了纹理和语义特征之间的关联,以及时序特征对检测性能提升的影响。为了解决上述问题,文中提出了一种用于深度伪造检测的、具有多样化特征的自适应网络(AdfNet),它通过提取语义特征、纹理特征和时序特征帮助分类器判断真伪;探索了自适应纹理噪声提取机制(ATNEM),通过未池化的特征映射与基于频域的通道注意力机制,灵活捕获非固定频段的噪声残差;设计了深层语义分析指导策略(DSAGS),通过空间注意力机制突出篡改痕迹,并引导特征提取器关注焦点区域的深层特征;研究了多尺度时序特征处理方法(MTFPM),利用时序注意力机制给不同视频帧分配权重,捕获被篡改视频中时间序列的差异。实验结果表明,所提出的网络在FaceForensics++(FF++)数据集HQ模式中的ACC值为97.41%,相比当前主流网络有较为明显的性能提升;并且在FF++数据集上保持AUC值为99.80%的同时,在CelebDF上AUC值可达到76.41%,具有较强的泛化性。

The harm caused by video tampering has been endangering people’s lives,which makes deep forgery detection technology gradually obtain widespread attention and development.However,current detection methods could not effectively capture noisy residuals due to the use of inflexible constraints.In addition,they ignore the correlation between texture and semantic features and the impact of temporal features on detection performance improvement.To solve these problems,this paper proposed an adaptive network(AdfNet)with diverse features for deep forgery detection.It helps the classifier to judge authenticity by extracting semantic features,texture features and temporal features.The paper explored the adaptive texture noise extraction(ATNEM)mechanism,and flexibly captured the noise residuals in non-fixed frequency bands through unpooled feature mapping and frequency-based channel attention mechanism.The deep semantic analysis guidance strategy(DSAGS)was designed to highlight the tampering traces through spatial attention mechanism,and guide the feature extractor to focus on the deep features of the focus region.The paper studied multi-scale temporal feature processing(MTFPM),and used temporal attention mechanism to assign weights to different video frames and capture the difference of time series in tampered videos.The experimental results show that the ACC score of the proposed network in the HQ mode of FaceForensics++(FF++)dataset is 97.41%,which is significantly better than that of the existing mainstream algorithms.Moreover,while maintaining the AUC value of 99.80%on the FF++dataset,the AUC value can reach 76.41%on Celeb-DF,reflecting strong generalization.