基于混合防御架构的深度伪造检测与防御系统

Deepfake Detection and Defense System Based on Hybrid Defense Architecture

摘要: 深度伪造(Deepfake)技术的快速演进使高质量伪造内容泛滥，对个人隐私、舆论安全与社会治理构成严峻威胁。现有防御方法普遍存在检测手段单一、主动与被动防御割裂、跨数据集泛化能力不足等问题。为此，文章提出一种“主动防御 + 被动检测”的混合防御架构。在主动防御方面，提出梯度引导的区域自适应对抗扰动生成方法，结合人脸显著性图与梯度信息实现感知不可见的精准扰动分配，并设计基于DWT-DCT-SVD级联变换与K-means盲判决的鲁棒盲水印方案，支持“在线溯源 + 本地验真”双模式可信链。在被动检测方面，构建多源证据融合检测框架：以改进的双分支注意力增强ResNet-50为核心检测器，融合频域–纹理–几何传统特征分析层与视觉大模型语义推理层，并引入基于Dempster-Shafer证据理论的自适应融合决策机制替代固定权重方案，实现多源证据的不确定性建模与冲突消解。在FaceForensics++数据集上的实验表明，该系统性能良好，面对不同类型伪造Face2Face、FaceSwap、NeuralTextures检测准确率分别达到97.12%、95.58%和89.44%，优于现有主流方法。

Abstract: The rapid evolution of Deepfake technology has led to the proliferation of high-quality forged content, posing a severe threat to personal privacy, public opinion security, and social governance. Existing defense methods generally have problems, such as single detection means, the separation of active and passive defenses, and insufficient cross-dataset generalization ability. Therefore, this paper proposes a hybrid defense architecture of “active defense + passive detection”. In terms of active defense, a method for generating gradient-guided region-adaptive adversarial perturbations is proposed. By combining face saliency maps and gradient information, precise perturbation allocation that is imperceptible to perception is achieved. A robust blind watermarking scheme based on the cascade transformation of DWT-DCT-SVD and K-means blind decision is designed to support the dual-mode trusted chain of “online traceability + local verification”. In terms of passive detection, a multi-source evidence fusion detection framework is constructed: an improved dual-branch attention-enhanced ResNet-50 is used as the core detector, which integrates the traditional feature analysis layers of frequency domain-texture-geometry and the semantic reasoning layer of large vision models. An adaptive fusion decision mechanism based on Dempster-Shafer evidence theory is introduced to replace the fixed weight scheme, realizing the uncertainty modeling and conflict resolution of multi-source evidence. Experiments on the FaceForensics++ dataset show that the system has good performance. The detection accuracies for different types of forgeries, Face2Face, FaceSwap, and NeuralTextures, reach 97.12%, 95.58%, and 89.44%, respectively, which are better than existing mainstream methods.