Scholarlink

<p>Multimedia forgeries are more realistic due to quick generative model development, complex detection algorithms that struggle with domain shifts, temporal inconsistencies, adversarial noise, and privacy constraints. A semantic disentanglement network for cross-domain generalization, a memory-guided temporal transformer for long-range video analysis, a perturbation-informed optimization strategy for adversarial robustness, a federated calibration mechanism with adaptive noise shaping for privacy-preserving learning, and a multidimensional evaluation matrix for unified evaluation address these limitations. Space localization, temporal coherence, perturbation robustness, and cross-domain accuracy across image and video benchmarks improve dramatically with the approach &amp; process. These results show the system’s suitability for forensic, security, and large-scale content-authentications. All in all, these models improve accuracy of forgery detection across modalities, adding up to 35% robustness against adversarial attacks while still maintaining performance within federated environments with less than 2.5% performance degradation.This lays a scalable, privacy-aware foundation for trustworthy multimedia forensics with cross-domain deployment capability and principled validation in process.</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

Generalizable and privacy-preserving multimedia forgery detection via semantically disentangled and temporally-aware deep learning architectures

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Shital Jadhav,
  • Mahip Bartere

摘要

Multimedia forgeries are more realistic due to quick generative model development, complex detection algorithms that struggle with domain shifts, temporal inconsistencies, adversarial noise, and privacy constraints. A semantic disentanglement network for cross-domain generalization, a memory-guided temporal transformer for long-range video analysis, a perturbation-informed optimization strategy for adversarial robustness, a federated calibration mechanism with adaptive noise shaping for privacy-preserving learning, and a multidimensional evaluation matrix for unified evaluation address these limitations. Space localization, temporal coherence, perturbation robustness, and cross-domain accuracy across image and video benchmarks improve dramatically with the approach & process. These results show the system’s suitability for forensic, security, and large-scale content-authentications. All in all, these models improve accuracy of forgery detection across modalities, adding up to 35% robustness against adversarial attacks while still maintaining performance within federated environments with less than 2.5% performance degradation.This lays a scalable, privacy-aware foundation for trustworthy multimedia forensics with cross-domain deployment capability and principled validation in process.