DefakeHop: A Light-Weight High-Performance Deepfake Detector

View on arXiv ← Back to list

Authors: Hong-Shuo Chen, Mozhdeh Rouhsedaghat, Hamza Ghani, Shuowen Hu, Suya You, C. -C. Jay Kuo

Published: 2021-03-11 20:01:30+00:00

AI Summary

DefakeHop is a lightweight deepfake detection method that uses successive subspace learning (SSL) to extract features from face images. It achieves state-of-the-art performance with a small model size, outperforming other methods on several benchmark datasets.

Abstract

A light-weight high-performance Deepfake detection method, called DefakeHop, is proposed in this work. State-of-the-art Deepfake detection methods are built upon deep neural networks. DefakeHop extracts features automatically using the successive subspace learning (SSL) principle from various parts of face images. The features are extracted by c/w Saab transform and further processed by our feature distillation module using spatial dimension reduction and soft classification for each channel to get a more concise description of the face. Extensive experiments are conducted to demonstrate the effectiveness of the proposed DefakeHop method. With a small model size of 42,845 parameters, DefakeHop achieves state-of-the-art performance with the area under the ROC curve (AUC) of 100%, 94.95%, and 90.56% on UADFV, Celeb-DF v1 and Celeb-DF v2 datasets, respectively.

Key findings

DefakeHop achieves state-of-the-art AUC scores (100%, 94.95%, and 90.56% on UADFV, Celeb-DF v1, and Celeb-DF v2, respectively) with only 42,845 parameters. Its performance is competitive even with significantly fewer training samples than other methods.

Approach

DefakeHop uses the successive subspace learning (SSL) principle with a c/w Saab transform to extract features from face image patches. A feature distillation module reduces dimensionality and incorporates soft classification, and an ensemble classification combines results from different facial regions and frames.

Datasets

UADFV, FaceForensics++, Celeb-DF v1, Celeb-DF v2

Model(s)

PixelHop++, c/w Saab transform, XGBoost

Author countries

USA

← Previous