基于困难感知元学习的跨域人脸伪造检测  

Difficulty-Aware Meta-Learning for Cross-Domain Face Forgery Detection

作　　者：金世辰 谭晓阳[1] JIN Shichen;TAN Xiaoyang(College of Computer Science and Technology,Nanjing University of Aeronautics&Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106

出　　处：《南京航空航天大学学报(自然科学版)》2025年第2期371-377,共7页Journal of Nanjing University of Aeronautics & Astronautics

基　　金：国家自然科学基金(6247072715)。

摘　　要：随着面部伪造技术的快速迭代,能够应对未见过的伪造方法的鲁棒检测机制需求变得日益重要。然而,当前的方法主要针对特定的伪造技术设计,这在应对更广泛的检测挑战时存在局限性。为了解决这些问题,本文提出了一种用于跨域人脸伪造检测的难度感知元学习(Difficulty⁃aware meta⁃learning,DAML)方法。在元训练阶段,本文方法利用与伪造图像无关的元学习(Model⁃agnostic meta⁃learning,MAML)方法来训练模型。通过利用目标域中的少量数据,可以调整参数以适应新任务。为了解决与模型无关的元学习方法中的不稳定训练问题,本文引入了一种难度感知机制,在训练阶段动态调整不同任务的学习权重。在多个公开的基准数据集上进行了广泛的实验,实验结果表明,本文方法优于RECCE、Xception、RFM等方法,在适应未见过的目标域方面表现更好。With the rapid iteration of facial forgery techniques,robust detection mechanisms that can handle unseen forgery methods are increasingly crucial.However,current approaches are primarily tailored to specific forgery techniques,posing limitations in addressing this broader detection challenge.To address these issues,this paper proposes a difficulty⁃aware meta⁃learning(DAML)method for cross⁃domain face forgery detection.During the meta⁃training phase,our method utilizes a model⁃agnostic meta⁃learning(MAML)approach,using historical forgery images to optimize the meta⁃leaner.By leveraging a small amount of data in the target domain,we can adjust the parameters to fit new tasks.To tackle the issue of unstable training in model⁃agnostic meta⁃learning methods,we introduce a difficulty⁃aware mechanism to dynamically adjust the learning weights for different tasks during the training phase.We conduct extensive experiments on several publicly available benchmark datasets.The experimental results demonstrate that our method outperforms several methods,such as:RECCE,Xception,RFM,etc.,achieving better adaptability to unseen target domains.

关 键 词：人脸伪造检测 元学习 跨领域 动态调整学习权重 泛化性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]