A method and system for detecting deepfake facial images

By constructing a differential loss function and noise-injected reprojection in the latent space, a face image detection model is trained, which solves the problem of insufficient generalization and robustness of existing models and achieves efficient detection under various forgery methods and noisy environments.

CN121600582BActive Publication Date: 2026-05-26ZHEJIANG GONGSHANG UNIVERSITY
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
ZHEJIANG GONGSHANG UNIVERSITY
Filing Date
2026-01-29
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing methods for detecting deepfake faces suffer from insufficient generalization and robustness due to the large gap between training and testing data, especially when there are few forgery clues, resulting in a significant decrease in detection accuracy.

Method used

A differentiated loss function is constructed. Noise is injected into the original image features in the latent space through a latent diffusion model and then reprojected. The face image detection model is trained using an encoder, a noise injection module, a latent space reprojection branch, and a classifier. This allows real face images and fake face images to form reprojection distances of different sizes in the latent space. A differentiated loss function is then constructed to constrain the reprojection process.

Benefits of technology

It improves the generalization and robustness of the face image detection model in cross-dataset testing and various forgery methods, and provides efficient detection capabilities in noisy environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121600582B_ABST
    Figure CN121600582B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for deepfake detection of face images, belonging to the field of deepfake detection technology. A training set is constructed. When the face image to be detected belongs to a real face image in the training set, the loss function is the minimum distance between the original image features and the real features. When the face image to be detected belongs to a fake face image in the training set, the loss function is the maximum difference between the outer margin of the fake face image and the distance between the original image features and the fake features, determined by the distance between the original image features and the fake features. A binary cross-entropy loss is constructed based on the difference between the labels and the detection results. A face image detection model for distinguishing between real and fake face images is trained based on the loss function and the binary cross-entropy loss. This method improves the detection accuracy, generalization, and robustness of the face image detection model by injecting noise into the latent space and using differential reprojection distance to distinguish between real and fake samples.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Face forgery detection method based on multi-domain clue reconstruction

    CN120088833A