FakePolisher: Making DeepFakes More Detection-Evasive by Shallow Reconstruction

View on arXiv ← Back to list

Authors: Yihao Huang, Felix Juefei-Xu, Run Wang, Qing Guo, Lei Ma, Xiaofei Xie, Jianwen Li, Weikai Miao, Yang Liu, Geguang Pu

Published: 2020-06-13 01:48:15+00:00

AI Summary

This paper introduces FakePolisher, a post-processing method that reduces artifacts in GAN-generated fake images by performing shallow reconstruction using a learned linear dictionary. This significantly decreases the accuracy of state-of-the-art deepfake detection methods, highlighting the reliance of current detectors on these artifacts.

Abstract

At this moment, GAN-based image generation methods are still imperfect, whose upsampling design has limitations in leaving some certain artifact patterns in the synthesized image. Such artifact patterns can be easily exploited (by recent methods) for difference detection of real and GAN-synthesized images. However, the existing detection methods put much emphasis on the artifact patterns, which can become futile if such artifact patterns were reduced. Towards reducing the artifacts in the synthesized images, in this paper, we devise a simple yet powerful approach termed FakePolisher that performs shallow reconstruction of fake images through a learned linear dictionary, intending to effectively and efficiently reduce the artifacts introduced during image synthesis. The comprehensive evaluation on 3 state-of-the-art DeepFake detection methods and fake images generated by 16 popular GAN-based fake image generation techniques, demonstrates the effectiveness of our technique.Overall, through reducing artifact patterns, our technique significantly reduces the accuracy of the 3 state-of-the-art fake image detection methods, i.e., 47% on average and up to 93% in the worst case.

Key findings

FakePolisher significantly reduces the accuracy of three state-of-the-art deepfake detection methods (47% on average, up to 93% in the worst case). The results demonstrate that current detection methods heavily rely on artifacts introduced during image synthesis. The reconstructed images maintain high similarity to their original fake counterparts.

Approach

FakePolisher trains a dictionary model on real images to capture their patterns. It then projects fake images onto this learned subspace using linear projection or sparse coding to reconstruct a 'fake-free' version of the image, reducing detectable artifacts.

Datasets

CelebA, LSUN, FFHQ; fake images generated by 16 GAN-based methods (including ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, CRN, IMLE, SITD, SAN, DeepFakes, StyleGAN2, and Which-faceisreal).

Model(s)

PCA and K-SVD dictionary learning models are used for reconstruction. The paper evaluates the effectiveness of the approach against three state-of-the-art deepfake detection methods: GANFingerprint, CNNDetector, and DCTA.

Author countries

China, USA, Singapore, Japan

← Previous