| 摘要: |
| 目的 针对现有人脸伪造检测方法泛化性不足、鲁棒性较差的问题,提出了一种基于对抗训练的改进人脸伪
造检测方法( ATNet) 。 方法 ATNet 通过动态合成对抗伪造样本扩大样本空间,强化模型对不同伪造算法生成的伪
造样本“ 敏感度” ,避免训练样本过采样导致的模型过拟合问题;替换特定地标点标识的伪造区域以提高模型学习
不同伪造特征的能力;采用高频噪声和低频纹理并行学习策略,融合多层次卷积特征,促使模型捕捉更具综合性的
伪造线索,使其更有效的鉴别伪造图像。 结果 随着有效的对抗样本生成,ATNet 可以学习到更为本质的伪造特征,
相对于当前较为优秀方法如 Xception,F3Net,Face X-ray 等在检测精度和泛化性上均有不同程度地提升,跨数据集
测试结果表明该模型具备优秀的泛化性能;使用降维算法进行可视化可以直观判断出 ATNet 对于深度伪造图像的
检测能力。 结论 基于改进对抗训练的人脸检测方法可以有效提升检测精度和泛化性,强化模型对于多种伪造特征
的“ 敏感度” ,在多个数据集上的实验结果表明 ATNet 是简单有效的。 |
| 关键词: 深度伪造 人脸伪造检测 对抗训练 数据增强 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Improved Face Forgery Detection Method Based on Adversarial Training |
|
ZHANG Kai1 FAN Zhixian2
|
|
1. School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
2. School of Artificial Intelligence Anhui University of Science and Technology Anhui Huainan 232001 China
|
| Abstract: |
| Objective Aiming at the problems of insufficient generalization and poor robustness of existing face forgery
detection methods an improved face forgery detection method based on adversarial training ATNet is proposed.
Methods ATNet expands the sample space by dynamically synthesizing adversarial forgery samples to enhance the model?? s
??sensitivity to forgery samples generated by different forgery algorithms and avoid the over-fitting of the model caused by
over-sampling of training samples. By replacing the forgery regions marked by specific landmark points the model?? s
ability to learn different forgery features is improved. A parallel learning strategy of high-frequency noise and low-
frequency textures is adopted to fuse multi-level convolutional features which enables the model to capture more
comprehensive forgery clues and more effectively identify forged images. Results With the effective generation of
adversarial samples ATNet can learn more essential forgery features. Compared with current excellent methods such as
Xception F3Net and Face X-ray ATNet has varying degrees of improvement in detection accuracy and generalization.
The cross-dataset test results show that the model has excellent generalization performance. Visualization using the
dimensionality reduction algorithm can intuitively determine ATNet ?? s ability to detect deep-forged images. Conclusion The face forgery detection method based on improved adversarial training can effectively improve the detection
accuracy and generalization and strengthen the model?? s ??sensitivity to various forgery features. The experimental results
on multiple datasets show that ATNet is simple and effective. |
| Key words: deepfakes face forgery detection adversarial training data augmentation |
