Scholarlink

<p>With rapid advances in Deep Learning (DL) algorithms and social media content, deepfakes have emerged as a potent tool for manipulating multimedia content to commit defamation, falsify information, and pose other security threats. Recent deepfake detection is challenging due to the intricate structure of deep learning models, high overlap between real and fake data, lower reliability, limited generalization, and poor interpretability and explainability of deepfake detection models. This paper presents a robust and reliable DeepFake Detection Network (DeepFakeDetNet) for videos using a Deep Convolutional Neural Network (DCNN) and Long Short-Term Memory (LSTM) networks to improve the generalization capability, interpretability, accuracy and minimize the computation intricacy of the model. The system considers the Image Texture and Shape Features (ITSF) that encompass the Gray level Co-occurrence Matrix (GLCM) feature to depict the spatial relationship in texture, a novel Extended Local Ternary Pattern (XLTP) to provide local texture patterns, and Histogram of Oriented Gradients (HOG) to depict the shape attributes of the deepfake images. Multiple Acoustic Features (MAFs) are used to depict the spectral, temporal, and phonetic attributes of the audio. Furthermore, an Improved Starfish Optimization Algorithm (ISOA) is employed for feature selection, focusing on prominent features to reduce the system’s computational complexity. The effectiveness of the proposed deepfake detection system is evaluated on the FakeAVCeleb dataset. The proposed ITSF + MAF+SOA offers better generalization capability, superior spectral-temporal depiction of multimodal deepfake modalities, and minimization in computational intricacy of the deepfake detection system. The ITSF + MAF+ISO achieves an improved accuracy of 97.80%, a recall of 98.29%, a precision of 97.10%, and an F1-score of 97.70% compared to the traditional state of the art.</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

Deepfake detection using deep convolutional neural network and long short-term memory

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Sujata Bahadure,
  • Vanita Mane

摘要

With rapid advances in Deep Learning (DL) algorithms and social media content, deepfakes have emerged as a potent tool for manipulating multimedia content to commit defamation, falsify information, and pose other security threats. Recent deepfake detection is challenging due to the intricate structure of deep learning models, high overlap between real and fake data, lower reliability, limited generalization, and poor interpretability and explainability of deepfake detection models. This paper presents a robust and reliable DeepFake Detection Network (DeepFakeDetNet) for videos using a Deep Convolutional Neural Network (DCNN) and Long Short-Term Memory (LSTM) networks to improve the generalization capability, interpretability, accuracy and minimize the computation intricacy of the model. The system considers the Image Texture and Shape Features (ITSF) that encompass the Gray level Co-occurrence Matrix (GLCM) feature to depict the spatial relationship in texture, a novel Extended Local Ternary Pattern (XLTP) to provide local texture patterns, and Histogram of Oriented Gradients (HOG) to depict the shape attributes of the deepfake images. Multiple Acoustic Features (MAFs) are used to depict the spectral, temporal, and phonetic attributes of the audio. Furthermore, an Improved Starfish Optimization Algorithm (ISOA) is employed for feature selection, focusing on prominent features to reduce the system’s computational complexity. The effectiveness of the proposed deepfake detection system is evaluated on the FakeAVCeleb dataset. The proposed ITSF + MAF+SOA offers better generalization capability, superior spectral-temporal depiction of multimodal deepfake modalities, and minimization in computational intricacy of the deepfake detection system. The ITSF + MAF+ISO achieves an improved accuracy of 97.80%, a recall of 98.29%, a precision of 97.10%, and an F1-score of 97.70% compared to the traditional state of the art.