Defending Deepfake via Texture Feature Perturbation

Authors: Xiao Zhang, Changfang Chen, Tianyi Wang

Published: 2025-08-24 11:53:35+00:00

Comment: Accepted to IEEE SMC 2025

AI Summary

This paper introduces a proactive Deepfake defense approach that embeds invisible perturbations into facial texture regions to disrupt Deepfake generation. The method leverages Local Binary Patterns (LBP) for initial texture extraction and a dual-model attention strategy to optimize these low-perceptual-saliency perturbations. Experiments on CelebA-HQ and LFW datasets demonstrate its effectiveness in causing visible distortions in Deepfake outputs under various attack models while maintaining high visual quality of the original protected images.

Abstract

The rapid development of Deepfake technology poses severe challenges to social trust and information security. While most existing detection methods primarily rely on passive analyses, due to unresolvable high-quality Deepfake contents, proactive defense has recently emerged by inserting invisible signals in advance of image editing. In this paper, we introduce a proactive Deepfake detection approach based on facial texture features. Since human eyes are more sensitive to perturbations in smooth regions, we invisibly insert perturbations within texture regions that have low perceptual saliency, applying localized perturbations to key texture regions while minimizing unwanted noise in non-textured areas. Our texture-guided perturbation framework first extracts preliminary texture features via Local Binary Patterns (LBP), and then introduces a dual-model attention strategy to generate and optimize texture perturbations. Experiments on CelebA-HQ and LFW datasets demonstrate the promising performance of our method in distorting Deepfake generation and producing obvious visual defects under multiple attack models, providing an efficient and scalable solution for proactive Deepfake detection.

Key findings

The proposed method achieves superior visual quality for protected images (higher PSNR, SSIM, lower LPIPS) compared to state-of-the-art defense techniques. It effectively distorts Deepfake generations, yielding high Defense Success Rates (DSR) and significant L2 norm distances between original and perturbed generated images across multiple Deepfake models. The framework also demonstrates strong generalization capability and robustness when evaluated on unseen datasets.

Approach

The authors propose a proactive Deepfake defense framework that inserts imperceptible perturbations into the texture regions of an image. It uses bilateral filtering and Local Binary Patterns (LBP) to extract preliminary texture features, then a dual-model attention strategy (involving ResNet50 and Vision Transformer with Grad-CAM) generates and optimizes these texture-guided perturbations. A multi-objective loss function combining L1 distance, L2 norm, and adversarial attention loss ensures visual fidelity while maximizing disruption to Deepfake generative models.

Datasets

CelebA-HQ, LFW

Model(s)

UNKNOWN

Author countries

China, Singapore

← Previous