Scholarlink

<p>The rise of DeepFake technology has raised significant concerns, as forged facial images pose threats to individuals, states, and societies. To assess the authenticity of videos, images, and audio, various DeepFake detection algorithms have been proposed to identify regions of facial forgery. However, challenges such as the complexity of multimodal information and biases in training datasets have resulted in a higher representation gap, limiting the performance of DeepFake detectors. To address these challenges, we propose a Multi-Collaborative Unsupervised Contrastive Learning (MCUCL) approach that investigates the factors causing cross-modal and intra-modal deepfakes. This method extracts features from three public deepfake datasets to generate non-demographic labels. We combine both cross-modal and intra-modal representations using Dempster-Shafer theory to detect unseen artifacts. Through this multi-collaborative contrastive learning, inter-frame correlation generates a binary output indicating real or forged data. Performance evaluation, conducted through complete sequence matching with multiple inferences, is carried out using various real-time DeepFake datasets. Performance evaluations on diverse real-world datasets show that MCUCL achieves 94.5% accuracy, 97.45% precision, and an AUC of 0.962, even on low-quality compressed videos. Extensive experiments validate the effectiveness of the proposed detection algorithm in identifying forged data compressed on social networks. Beyond technical performance, this proposed MCUCL offers a practical safeguard against manipulated content, helping protect individuals and societies from the harmful impacts of DeepFake.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-Collaborative Unsupervised Contrastive Learning for DeepFake Detection Amidst Unseen Artifact

  • 期刊论文

  • 发表日期：

    2025年11月6日

查看全文
  • Mary Sumitha Maria Michael,
  • Pandia Rajan Jeyaraj

摘要

The rise of DeepFake technology has raised significant concerns, as forged facial images pose threats to individuals, states, and societies. To assess the authenticity of videos, images, and audio, various DeepFake detection algorithms have been proposed to identify regions of facial forgery. However, challenges such as the complexity of multimodal information and biases in training datasets have resulted in a higher representation gap, limiting the performance of DeepFake detectors. To address these challenges, we propose a Multi-Collaborative Unsupervised Contrastive Learning (MCUCL) approach that investigates the factors causing cross-modal and intra-modal deepfakes. This method extracts features from three public deepfake datasets to generate non-demographic labels. We combine both cross-modal and intra-modal representations using Dempster-Shafer theory to detect unseen artifacts. Through this multi-collaborative contrastive learning, inter-frame correlation generates a binary output indicating real or forged data. Performance evaluation, conducted through complete sequence matching with multiple inferences, is carried out using various real-time DeepFake datasets. Performance evaluations on diverse real-world datasets show that MCUCL achieves 94.5% accuracy, 97.45% precision, and an AUC of 0.962, even on low-quality compressed videos. Extensive experiments validate the effectiveness of the proposed detection algorithm in identifying forged data compressed on social networks. Beyond technical performance, this proposed MCUCL offers a practical safeguard against manipulated content, helping protect individuals and societies from the harmful impacts of DeepFake.