Scholarlink

The exponential rise of deepfake videos and audios presents substantial threats to societal trust, individual privacy, and digital integrity. Deepfakes exploit advanced artificial intelligence to fabricate highly realistic visual and auditory content, resulting in misuse in misinformation, defamation, and fraud. This research presents a novel detection method which analyzes the dimensional and sequential features in the input vectors using ResNeXt-50 CNN and Long Short-Term Memory (LSTM). Even when applied to medium–resolution inputs, the model showcases strong performance. The model was evaluated on well-established datasets, including DFDC and FaceForensics++, showcasing its effectiveness in identifying manipulated content. Despite the promising results, issues such as dataset compression and the diversity of emerging deepfake techniques persist. Future enhancements could focus on high-resolution datasets, advanced augmentation methods, and real-time detection capabilities. Furthermore, incorporating interpretability and reliability into detection models would significantly aid forensic investigations and legal proceedings, fostering broader societal trust.

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

Deepfake Detection: Leveraging CNN-LSTM Architectures for Enhanced Spatial-Temporal Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ameya Khyati,
  • Shipra Shukla,
  • Ayushi Vigya,
  • Devansh Pratap Singh,
  • Diwakar Mishra

摘要

The exponential rise of deepfake videos and audios presents substantial threats to societal trust, individual privacy, and digital integrity. Deepfakes exploit advanced artificial intelligence to fabricate highly realistic visual and auditory content, resulting in misuse in misinformation, defamation, and fraud. This research presents a novel detection method which analyzes the dimensional and sequential features in the input vectors using ResNeXt-50 CNN and Long Short-Term Memory (LSTM). Even when applied to medium–resolution inputs, the model showcases strong performance. The model was evaluated on well-established datasets, including DFDC and FaceForensics++, showcasing its effectiveness in identifying manipulated content. Despite the promising results, issues such as dataset compression and the diversity of emerging deepfake techniques persist. Future enhancements could focus on high-resolution datasets, advanced augmentation methods, and real-time detection capabilities. Furthermore, incorporating interpretability and reliability into detection models would significantly aid forensic investigations and legal proceedings, fostering broader societal trust.