Training Strategies and Data Augmentations in CNN-based DeepFake Video Detection

View on arXiv ← Back to list

Authors: Luca Bondi, Edoardo Daniele Cannas, Paolo Bestagini, Stefano Tubaro

Published: 2020-11-16 08:50:56+00:00

AI Summary

This paper investigates the impact of training strategies and data augmentation techniques on the performance of CNN-based deepfake video detectors. The authors analyze intra-dataset and cross-dataset detection performance using various training methods and augmentations.

Abstract

The fast and continuous growth in number and quality of deepfake videos calls for the development of reliable detection systems capable of automatically warning users on social media and on the Internet about the potential untruthfulness of such contents. While algorithms, software, and smartphone apps are getting better every day in generating manipulated videos and swapping faces, the accuracy of automated systems for face forgery detection in videos is still quite limited and generally biased toward the dataset used to design and train a specific detection system. In this paper we analyze how different training strategies and data augmentation techniques affect CNN-based deepfake detectors when training and testing on the same dataset or across different datasets.

Key findings

Data augmentation significantly improves cross-dataset generalization, with some augmentations more beneficial than others. Triplet loss helps improve performance with limited training data, but data augmentation with BCE loss performs better on larger datasets. The largest dataset (DFDC) generally generalizes better across datasets.

Approach

The authors utilize an EfficientNetB4 architecture trained with Binary Cross Entropy (BCE) loss and triplet loss, evaluating the impact of different training strategies and data augmentation techniques (horizontal flip, brightness/contrast, HSV changes, ISO noise, Gaussian noise, downscaling, JPEG compression) on deepfake detection accuracy both within and across different datasets.

Datasets

FaceForensics++, DeepFake Detection Challenge Dataset, CelebDF(v2)

Model(s)

EfficientNetB4

Author countries

Italy

← Previous