Face forgery detection method and system based on federal incremental learning

By employing a federated incremental learning framework and a difference-aware aggregation strategy, this study addresses the computational overhead and privacy issues of face forgery detection methods when facing new forgery techniques. It achieves continuous adaptation and stable recognition capabilities for the model, making it suitable for multi-source heterogeneous forgery threat scenarios.

CN122116490APending Publication Date: 2026-05-29GUANGZHOU UNIVERSITY
1 Cites 1 Cited by

Patent Information

Application Number
CN202610091740.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-23
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Existing face forgery detection methods require full retraining when faced with new forgery techniques, resulting in huge computational costs and difficulty in real-time adaptation. Centralized training leads to data privacy risks, lacks cross-data domain collaborative updates, and the model is prone to overfitting and catastrophic forgetting.

Method used

A federated incremental learning framework is adopted, which builds multiple clients and a global server for basic training and incremental training. Adversarial perturbation is used to generate fake alternative samples. Feature knowledge distillation and soft label distillation are combined, and a difference-aware aggregation strategy is used to update the global model, thereby achieving cross-task generalization and robustness.

Benefits of technology

This enables continuous adaptation and rapid updates of the model without uploading the original data, reducing storage overhead and privacy risks, mitigating the forgetting problem, and improving the ability and stability to identify new forgery types.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a face forgery detection method and system based on federal incremental learning, and the method comprises the following steps: constructing a federal incremental learning framework, each client being provided with a local face forgery detection model, and a global server being provided with a global face forgery detection model; performing cross entropy training on the local face forgery detection model in a basic training stage, generating a forgery substitution sample in an incremental training stage, taking the forgery substitution sample, a real face sample and a forgery face sample as training input, and uploading model parameters to a global server for aggregation after training; updating the global face forgery detection model based on a difference perception aggregation strategy; training confrontation disturbance based on the global face forgery detection model converged at the stage and a real face sample, and refreshing a disturbance pool until a preset termination condition is reached; and based on the trained global face forgery detection model, carrying out authenticity identification on a to-be-detected face image. According to the method, disastrous forgetting can be relieved, and the cross-task generalization ability and robustness can be improved.
Need to check novelty before this filing date? Find Prior Art