Landmark Breaker: Obstructing DeepFake By Disturbing Landmark Extraction

View on arXiv ← Back to list

Authors: Pu Sun, Yuezun Li, Honggang Qi, Siwei Lyu

Published: 2021-02-01 12:27:08+00:00

AI Summary

This paper introduces Landmark Breaker, a novel method to prevent deepfake video generation by disrupting facial landmark extraction. It uses adversarial perturbations to mislead facial landmark extractors, thereby degrading the quality of the synthesized faces before the deepfake is even created.

Abstract

The recent development of Deep Neural Networks (DNN) has significantly increased the realism of AI-synthesized faces, with the most notable examples being the DeepFakes. The DeepFake technology can synthesize a face of target subject from a face of another subject, while retains the same face attributes. With the rapidly increased social media portals (Facebook, Instagram, etc), these realistic fake faces rapidly spread though the Internet, causing a broad negative impact to the society. In this paper, we describe Landmark Breaker, the first dedicated method to disrupt facial landmark extraction, and apply it to the obstruction of the generation of DeepFake videos.Our motivation is that disrupting the facial landmark extraction can affect the alignment of input face so as to degrade the DeepFake quality. Our method is achieved using adversarial perturbations. Compared to the detection methods that only work after DeepFake generation, Landmark Breaker goes one step ahead to prevent DeepFake generation. The experiments are conducted on three state-of-the-art facial landmark extractors using the recent Celeb-DF dataset.

Key findings

Landmark Breaker effectively increases the Normalized Mean Error (NME) of landmark detection and decreases the Structural Similarity (SSIM) of synthesized faces, demonstrating its ability to disrupt deepfake generation. However, the method shows limited transferability across different landmark extractors and is more robust to image compression than baseline methods.

Approach

Landmark Breaker adds adversarial perturbations to input images to disrupt facial landmark extraction. This is achieved by introducing a new loss function that maximizes the error between predicted and original heatmaps, optimized using the momentum iterative fast gradient sign method (MI-FGSM). The disrupted landmarks then hinder the alignment process crucial for deepfake generation.

Datasets

Celeb-DF dataset

Model(s)

FAN, HRNet, and AVS-SAN (for landmark extraction); A pre-trained DeepFake generation model from the Celeb-DF dataset is used for evaluation.

Author countries

China, USA

← Previous