Exploring Active Data Selection Strategies for Continuous Training in Deepfake Detection

View on arXiv ← Back to list

Authors: Yoshihiko Furuhashi, Junichi Yamagishi, Xin Wang, Huy H. Nguyen, Isao Echizen

Published: 2025-02-11 05:35:36+00:00

AI Summary

This paper proposes an active data selection method for continuously training deepfake detection models. The method uses model confidence scores to select a small subset of data from a large pool, significantly improving detection performance with minimal data.

Abstract

In deepfake detection, it is essential to maintain high performance by adjusting the parameters of the detector as new deepfake methods emerge. In this paper, we propose a method to automatically and actively select the small amount of additional data required for the continuous training of deepfake detection models in situations where deepfake detection models are regularly updated. The proposed method automatically selects new training data from a textit{redundant} pool set containing a large number of images generated by new deepfake methods and real images, using the confidence score of the deepfake detection model as a metric. Experimental results show that the deepfake detection model, continuously trained with a small amount of additional data automatically selected and added to the original training set, significantly and efficiently improved the detection performance, achieving an EER of 2.5% with only 15% of the amount of data in the pool set.

Key findings

Continuous training with actively selected data (15% of the pool set) significantly improved deepfake detection performance, reducing the Equal Error Rate (EER) from 22.5% to 2.5%. Active selection outperformed random selection in later iterations.

Approach

The authors propose a method that selects data for continuous training based on the confidence scores of a deepfake detection model. Low confidence scores indicate potential misclassifications and are used to select new training data from a pool of images, improving the model's performance incrementally.

Datasets

ForgeryNet (starter master set), FF++, Google DFD, YouTube DF, KoDF, Stable Diffusion 2.1, VoxCeleb, FFHQ (pool set)

Model(s)

EfficientNet V2-M (backbone) with a binary classification head. Trained using AdamW optimizer.

Author countries

Japan

← Previous