BiHPF: Bilateral High-Pass Filters for Robust Deepfake Detection

View on arXiv ← Back to list

Authors: Yonghyun Jeong, Doyeon Kim, Seungjai Min, Seongho Joe, Youngjune Gwon, Jongwon Choi

Published: 2021-08-16 07:56:45+00:00

AI Summary

This paper proposes Bilateral High-Pass Filters (BiHPF) for robust deepfake detection. BiHPF amplifies frequency-level artifacts present in synthesized images, improving detection accuracy even in unseen domains. The method outperforms state-of-the-art techniques on various datasets.

Abstract

The advancement in numerous generative models has a two-fold effect: a simple and easy generation of realistic synthesized images, but also an increased risk of malicious abuse of those images. Thus, it is important to develop a generalized detector for synthesized images of any GAN model or object category, including those unseen during the training phase. However, the conventional methods heavily depend on the training settings, which cause a dramatic decline in performance when tested with unknown domains. To resolve the issue and obtain a generalized detection ability, we propose Bilateral High-Pass Filters (BiHPF), which amplify the effect of the frequency-level artifacts that are known to be found in the synthesized images of generative models. Numerous experimental results validate that our method outperforms other state-of-the-art methods, even when tested with unseen domains.

Key findings

The proposed BiHPF method significantly outperforms existing state-of-the-art deepfake detection methods. It shows strong cross-domain performance, maintaining high accuracy even when tested on unseen categories, color variations, and GAN models. Ablation studies confirm the effectiveness of both high-pass filter components.

Approach

The approach uses Bilateral High-Pass Filters (BiHPF) to enhance frequency-level artifacts in images. BiHPF consists of pixel-level and frequency-level high-pass filters that emphasize artifacts in the background and high-frequency components, respectively. A classification model then uses the filtered magnitude spectrum map as input.

Datasets

LSUN, FFHQ, and various generated images from ProGAN, StyleGAN, StyleGAN2, BigGAN, CycleGAN, and StarGAN. Datasets included multiple categories of subjects and color manipulations.

Model(s)

ResNet-50 (pre-trained on ImageNet)

Author countries

Korea

← Previous