| 摘要: |
| 目的 针对当前 Deepfake 检测侧重全局伪造特征,而局部纹理差异特征利用不足导致模型泛化性能差的问
题,提出一种基于局部纹理差异特征增强的 Deepfake 检测模型,通过挖掘伪造图像内在的空间伪造模式,提高检测
的准确性和泛化性。 方法 模型首先通过中心差分卷积操作捕捉像素强度和像素梯度两种信息,从而获得更精确的
局部纹理差异信息,提高对伪造图像的敏感性。 其次,构建双层注意力模块,旨在利用空间注意力学习位置敏感的
权重信息,并通过通道注意力自适应调整通道重要性,定位重要纹理差异特征的位置,增强纹理差异特征的表示。
结果 在高质量和低质量的 FaceForensics++数据集上的实验,平均准确率分别达到了 97. 36%和 92. 37%,而 Celeb-
DF 数据集上的跨数据集实验获得了比当前先进的检测模型更好的泛化性,大量的消融实验表明了方法的有效性。
结论 实验表明:引入中心差分和双层注意力模块后模型能够更好地捕捉图像的纹理差异信息,适应不同场景和压
缩率的伪造检测,有效提高了 Deepfake 检测的准确性和泛化性。 |
| 关键词: Deepfake 检测 纹理差异 中心差分卷积 空间注意力 通道注意力 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Deepfake Detection Based on Local Texture Difference Feature Enhancement |
|
WEI Zhengzheng
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective Current Deepfake detection methods primarily focus on global forgery features leading to poor
generalization performance of the model due to insufficient utilization of local texture contrast features. To address this
issue a Deepfake detection model based on local texture difference feature enhancement was proposed aiming to improve
detection accuracy and generalization by exploring intrinsic spatial forgery patterns in forged images. Methods Firstly the
model captured both pixel intensity and pixel gradient by center difference convolution operation to obtain more accurate
local texture difference information and improve the sensitivity to forged images. Secondly a dual-layer attention module
was constructed aiming to use spatial attention to learn location-sensitive weighting information and adaptively adjust the
channel importance through channel attention to locate the position of important texture disparity features and enhance the
representation of texture disparity features. Results Experiments on high-quality and low-quality FaceForensics + +
datasets obtained average accuracies of 97. 36% and 92. 37% respectively while cross-dataset experiments on the Celeb-
DF dataset obtained better generalization performance than current state-of-the-art detection models. Extensive ablation
studies validate the effectiveness of the proposed method. Conclusion Experiments show that integrating center difference
convolution and a dual-layer attention module enables the model to better capture texture difference information in images
adapt to different scenarios and compression rates in forgery detection and effectively improve the accuracy and
generalization of Deepfake detection. |
| Key words: Deepfake detection texture difference center difference convolution spatial attention channel attention |
