Scholarlink

<p>The widespread availability of powerful generative models, such as Generative Adversarial Networks (GANs) and Autoencoders, has led to a dramatic surge in deepfake content production. These synthetically generated videos, often indistinguishable from authentic ones, pose a growing threat to digital trust and can facilitate misinformation, identity fraud, and social manipulation. While various deepfake detection models have been proposed, many struggle to maintain consistent performance across datasets and unseen video formats, revealing a clear research gap in generalizability and temporal feature modeling. In this proposed work, we introduce a novel deepfake detection framework that incorporates a customized Gated Recurrent Unit (GRU) architecture enhanced with a deep-fake-specific gating mechanism. This newly designed gate modifies the candidate hidden state update, enabling the network to more effectively capture temporal inconsistencies and facial dynamics in manipulated video sequences. The objective of this study is to design a deepfake detection system that not only performs with high accuracy but also generalizes well across varying datasets and video conditions. The proposed methodology employs a multi-stage pipeline. Vision Transformers (ViTs) are used initially to extract frame-level spatial features, followed by pretrained Convolutional Neural Networks (CNNs) such as MobileNetV2 to refine feature representations. These features are then given to a hybrid GRU architecture, combining both the standard and custom GRU unit, thereby enhancing sequential learning. The performance of the model is maximized by employing a Genetic Algorithm for optimizing GRU, number of layers and hyperparameter tuning. Benchmark datasets such as Celeb-DF V2 and FaceForensics + + have been utilized to train and test the system. Experimental results demonstrate that the proposed model outperforms existing state-of-the-art methods, achieving superior accuracy, precision, recall, F1-score, and ROC-AUC compared to all evaluated models. The proposed framework demonstrates strong generalization to unseen data and significantly reduces false positive rates. This provides a scalable and adaptive approach to deepfake detection by combining novel recurrent neural architecture with evolutionary optimization algorithms, demonstrating effectiveness on diverse datasets such as digital forensics, content verification system and policy enforcement platforms aimed at mitigating deepfake threats.</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

Attention-augmented hybrid framework with evolutionary optimization for robust deepfake detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月19日

查看全文
  • S. J. Shivaprakash,
  • Sabireen H,
  • Akshat Chauhan,
  • Abdul Quadir Md

摘要

The widespread availability of powerful generative models, such as Generative Adversarial Networks (GANs) and Autoencoders, has led to a dramatic surge in deepfake content production. These synthetically generated videos, often indistinguishable from authentic ones, pose a growing threat to digital trust and can facilitate misinformation, identity fraud, and social manipulation. While various deepfake detection models have been proposed, many struggle to maintain consistent performance across datasets and unseen video formats, revealing a clear research gap in generalizability and temporal feature modeling. In this proposed work, we introduce a novel deepfake detection framework that incorporates a customized Gated Recurrent Unit (GRU) architecture enhanced with a deep-fake-specific gating mechanism. This newly designed gate modifies the candidate hidden state update, enabling the network to more effectively capture temporal inconsistencies and facial dynamics in manipulated video sequences. The objective of this study is to design a deepfake detection system that not only performs with high accuracy but also generalizes well across varying datasets and video conditions. The proposed methodology employs a multi-stage pipeline. Vision Transformers (ViTs) are used initially to extract frame-level spatial features, followed by pretrained Convolutional Neural Networks (CNNs) such as MobileNetV2 to refine feature representations. These features are then given to a hybrid GRU architecture, combining both the standard and custom GRU unit, thereby enhancing sequential learning. The performance of the model is maximized by employing a Genetic Algorithm for optimizing GRU, number of layers and hyperparameter tuning. Benchmark datasets such as Celeb-DF V2 and FaceForensics + + have been utilized to train and test the system. Experimental results demonstrate that the proposed model outperforms existing state-of-the-art methods, achieving superior accuracy, precision, recall, F1-score, and ROC-AUC compared to all evaluated models. The proposed framework demonstrates strong generalization to unseen data and significantly reduces false positive rates. This provides a scalable and adaptive approach to deepfake detection by combining novel recurrent neural architecture with evolutionary optimization algorithms, demonstrating effectiveness on diverse datasets such as digital forensics, content verification system and policy enforcement platforms aimed at mitigating deepfake threats.