A Texture-Guided Adversarial Defense Framework Against Deepfake Generation

Deepfakes are an increasing danger to the reliability of digital media, with several detection algorithms developed to identify manipulated content. Although detection approaches continue to improve, proactive defenses that protect images by embedding imperceptible adversarial perturbations before distribution, thereby preventing deepfake generation at the source, offer a more robust complementary solution. However, existing proactive adversarial defenses against deepfake generators suffer from three key limitations: poor imperceptibility, limited texture-level targeting, and a lack of control over low-level structure perturbation. To address this, a novel patch-wise texture-aware adversarial attack is proposed as a proactive defense that disrupts deepfake generation. This is because adversarial images are generated through the addition of perturbations to the original facial images, so maintaining visual imperceptibility is a critical requirement; thus, our defense explicitly focuses on limiting perturbations to low-energy texture aspects that do not significantly affect human perception. The proposed method leverages tail-mode singular vectors extracted via patch-wise Singular Value Decomposition (SVD) to amplify low-energy texture components, i.e., producing adversarial perturbations that remain visually imperceptible while degrading the quality of generated deepfake images. Additionally, we introduce a variant that performs gradient projection into the SVD-tail subspace, ensuring that updates stay confined to subtle texture patterns and avoid high-frequency noise. The findings show that the proposed method is a stealthier attack that disrupts deepfake generations. Experimental results show that the proposed method achieves high attack success rates while preserving perceptual similarity, outperforming baselines in terms of imperceptibility and effectiveness. This work demonstrates a novel pathway toward proactive and imperceptible defenses against deepfake generation.