MUMBAI, India, Jan. 9 -- Intellectual Property India has published a patent application (202511112964 A) filed by Madan Mohan Malaviya University Of Technology, Gorakhpur, Uttar Pradesh, on Nov. 17, 2025, for 'attention-enhanced multi-stage framework for robust fake face image detection.'
Inventor(s) include Dr. Shantanu Shahi.
The application for the patent was published on Jan. 9, under issue no. 02/2026.
According to the abstract released by the Intellectual Property India: "The invention discloses with the rise of generative adversarial networks (GANs) and diffusion models, fake face images (deepfakes) have become increasingly realistic, threatening security, privacy, and social trust. While convolutional neural networks (CNNs) achieve strong performance in detecting manipulated faces, their limited receptive attention and vulnerability to example forgetting reduce robustness in real-world scenarios. Building upon prior two-stage attention-expanded detectors, this paper proposes an Attention-Enhanced Multi-Stage Detection Framework (AMDF) that fuses local, global, and frequency-domain features across three complementary learning stages. Unlike conventional approaches, AMDF integrates attention-expansion, frequency-aware residual modules, and a memory-preserving contrastive learning strategy to overcome feature forgetting. Extensive experiments on FF++, Celeb-DF, DFDC, HFF, and newly curated Diffusion-DeepFake (DDF) datasets demonstrate that AMDF achieves superior generalization, with 3-5% AUC improvement over state-of-the-art detectors, particularly on cross-dataset transfer tasks."
Disclaimer: Curated by HT Syndication.
