Uncovering Critical Features for Deepfake Detection through the Lottery Ticket Hypothesis

View on arXiv ← Back to list

Authors: Lisan Al Amin, Md. Ismail Hossain, Thanh Thi Nguyen, Tasnim Jahan, Mahbubul Islam, Faisal Quader

Published: 2025-07-21 13:58:24+00:00

AI Summary

This research explores the Lottery Ticket Hypothesis (LTH) for efficient deepfake detection, aiming to identify crucial features for recognizing deepfakes by pruning neural networks while maintaining accuracy. Experiments using MesoNet, CNN-5, and ResNet-18 on OpenForensic and FaceForensics++ datasets show that deepfake detection networks contain subnetworks that preserve performance even at high sparsity levels.

Abstract

Recent advances in deepfake technology have created increasingly convincing synthetic media that poses significant challenges to information integrity and social trust. While current detection methods show promise, their underlying mechanisms remain poorly understood, and the large sizes of their models make them challenging to deploy in resource-limited environments. This study investigates the application of the Lottery Ticket Hypothesis (LTH) to deepfake detection, aiming to identify the key features crucial for recognizing deepfakes. We examine how neural networks can be efficiently pruned while maintaining high detection accuracy. Through extensive experiments with MesoNet, CNN-5, and ResNet-18 architectures on the OpenForensic and FaceForensics++ datasets, we find that deepfake detection networks contain winning tickets, i.e., subnetworks, that preserve performance even at substantial sparsity levels. Our results indicate that MesoNet retains 56.2% accuracy at 80% sparsity on the OpenForensic dataset, with only 3,000 parameters, which is about 90% of its baseline accuracy (62.6%). The results also show that our proposed LTH-based iterative magnitude pruning approach consistently outperforms one-shot pruning methods. Using Grad-CAM visualization, we analyze how pruned networks maintain their focus on critical facial regions for deepfake detection. Additionally, we demonstrate the transferability of winning tickets across datasets, suggesting potential for efficient, deployable deepfake detection systems.

Key findings

Deepfake detection networks contain "winning tickets" (subnetworks) that maintain high accuracy even with 80% sparsity. The iterative pruning method consistently outperforms one-shot pruning. Grad-CAM visualizations show that pruned networks maintain focus on critical facial regions for deepfake detection.

Approach

The authors employ an iterative magnitude pruning approach based on the Lottery Ticket Hypothesis to identify and utilize subnetworks within existing deepfake detection models. This process involves iteratively pruning weights based on magnitude, retraining, and visualizing the impact using Grad-CAM. They compare this iterative method to a one-shot pruning approach.

Datasets

OpenForensic and FaceForensics++ datasets

Model(s)

MesoNet, CNN-5, ResNet-18, XceptionNet

Author countries

USA, Bangladesh, Australia

← Previous