A New Approach to Improve Learning-based Deepfake Detection in Realistic Conditions

View on arXiv ← Back to list

Authors: Yuhang Lu, Touradj Ebrahimi

Published: 2022-03-22 15:16:54+00:00

AI Summary

This paper proposes a novel data augmentation scheme for improving the robustness of deepfake detectors to real-world image degradations like compression, noise, and enhancement. The augmentation, inspired by real-world image processing pipelines, enhances the generalization ability of deepfake detection models to unseen datasets and distortions.

Abstract

Deep convolutional neural networks have achieved exceptional results on multiple detection and recognition tasks. However, the performance of such detectors are often evaluated in public benchmarks under constrained and non-realistic situations. The impact of conventional distortions and processing operations found in imaging workflows such as compression, noise, and enhancement are not sufficiently studied. Currently, only a few researches have been done to improve the detector robustness to unseen perturbations. This paper proposes a more effective data augmentation scheme based on real-world image degradation process. This novel technique is deployed for deepfake detection tasks and has been evaluated by a more realistic assessment framework. Extensive experiments show that the proposed data augmentation scheme improves generalization ability to unpredictable data distortions and unseen datasets.

Key findings

The proposed data augmentation significantly improves the robustness of deepfake detectors against various realistic distortions and processing operations. The method maintains high performance on original unaltered data while significantly improving performance on data with added noise, compression, blurring, and other degradations. Cross-dataset experiments show improved generalization to unseen datasets.

Approach

The authors address the problem of deepfake detection robustness by introducing a data augmentation technique that simulates real-world image degradation processes. This involves stochastically applying image enhancement, blurring, Gaussian noise, and JPEG compression to the training data, creating a more realistic and robust training environment.

Datasets

FaceForensics++ (FFpp) and Celeb-DFv2 datasets.

Model(s)

Capsule-Forensics and XceptionNet.

Author countries

Switzerland

← Previous