Automated Deepfake Detection

View on arXiv ← Back to list

Authors: Ping Liu, Yuewei Lin, Yang He, Yunchao Wei, Liangli Zhen, Joey Tianyi Zhou, Rick Siow Mong Goh, Jingen Liu

Published: 2021-06-20 14:48:50+00:00

AI Summary

This paper introduces Automated Deepfake Detection (ADD), the first method to utilize Automated Machine Learning (AutoML) for deepfake detection. ADD adaptively searches for a neural architecture and incorporates a strategy to estimate potential manipulation regions, improving generalization and reducing reliance on prior knowledge.

Abstract

In this paper, we propose to utilize Automated Machine Learning to adaptively search a neural architecture for deepfake detection. This is the first time to employ automated machine learning for deepfake detection. Based on our explored search space, our proposed method achieves competitive prediction accuracy compared to previous methods. To improve the generalizability of our method, especially when training data and testing data are manipulated by different methods, we propose a simple yet effective strategy in our network learning process: making it to estimate potential manipulation regions besides predicting the real/fake labels. Unlike previous works manually design neural networks, our method can relieve us from the high labor cost in network construction. More than that, compared to previous works, our method depends much less on prior knowledge, e.g., which manipulation method is utilized or where exactly the fake image is manipulated. Extensive experimental results on two benchmark datasets demonstrate the effectiveness of our proposed method for deepfake detection.

Key findings

ADD achieves competitive accuracy compared to existing methods, especially in cross-dataset evaluations where training and testing data are manipulated by different methods. The automatic architecture search yields a unique and effective network structure. The incorporation of potential manipulation region localization significantly boosts generalization performance.

Approach

ADD employs AutoML to automatically search for an optimal neural architecture for deepfake detection. It simultaneously learns to classify real/fake images and localize potential manipulation regions, improving generalization across different deepfake creation methods. The architecture search uses a gradient-based method, and the model is trained with a combined cross-entropy and mean squared error loss.

Datasets

FaceForensics++ (FF++) and Celeb-DF

Model(s)

A neural network architecture automatically searched using AutoML, composed of hierarchically stacked normal and reduction cells. The cells consist of operations selected from a search space including various convolution and pooling operations.

Author countries

Singapore, Australia, USA

← Previous