Implicit Identity Leakage: The Stumbling Block to Improving Deepfake Detection Generalization

View on arXiv ← Back to list

Authors: Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, Zheng Ge

Published: 2022-10-26 04:02:29+00:00

AI Summary

This paper identifies "Implicit Identity Leakage" as a major obstacle to generalizing deepfake detection models, where models unintentionally learn identity representations instead of focusing on forgery artifacts. To address this, the authors propose an ID-unaware model that uses an Artifact Detection Module to focus on local image features, improving cross-dataset generalization.

Abstract

In this paper, we analyse the generalization ability of binary classifiers for the task of deepfake detection. We find that the stumbling block to their generalization is caused by the unexpected learned identity representation on images. Termed as the Implicit Identity Leakage, this phenomenon has been qualitatively and quantitatively verified among various DNNs. Furthermore, based on such understanding, we propose a simple yet effective method named the ID-unaware Deepfake Detection Model to reduce the influence of this phenomenon. Extensive experimental results demonstrate that our method outperforms the state-of-the-art in both in-dataset and cross-dataset evaluation. The code is available at https://github.com/megvii-research/CADDM.

Key findings

The proposed ID-unaware model significantly outperforms state-of-the-art methods in both in-dataset and cross-dataset evaluations, demonstrating the effectiveness of mitigating Implicit Identity Leakage. The model also shows improved robustness to image perturbations.

Approach

The authors propose an ID-unaware Deepfake Detection Model that incorporates an Artifact Detection Module. This module uses multi-scale anchors to detect local artifact areas, reducing reliance on global identity information for classification. A Multi-scale Facial Swap method is used to generate training data with accurate artifact location annotations.

Datasets

FaceForensics++ (FF++), Celeb-DF, DFDC-V2

Model(s)

ResNet-18, ResNet-34, ResNet-50, Xception, EfficientNet-b3, EfficientNet-b4. The main contribution uses these as backbones for the proposed ID-unaware model.

Author countries

UNKNOWN

← Previous