Generalizable Deepfake Detection Framework Using Hybrid Convolution-Based EfficientNetB7 with Attention Mechanism - Deepfake video detection

Jul 30, 2026

Upload by Jiven

Highly realistic deepfake videos generated by recent deepfake technologies can spread misinformation with serious political and societal consequences, motivating automated detection methods. A hybrid deep learning approach is proposed to distinguish real from fake videos by leveraging facial landmark features and eye blinks extracted from video frames. Facial texture and shape features plus blink counts are fused with optimally tuned weights using Updated Random Parameter-based Fennec fox optimization, then classified via a Hybrid 2D-3D EfficientNetB7 with attention mechanism. Experiments validate strong performance with accuracies of 93.66% on dataset 1 and 94.38% on dataset 2.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Introduction

Deepfake background and risks

Deepfake categories and representation

Prior work and dataset limitations

Proposed approach

Video frame collection and feature extraction

Facial landmark and eye-blink based feature fusion

Optimized weighted fusion and HC-EB7AM classification

Experimental validation

Performance measures and results

Expert Systems With Applications 297 (2026) 129357 Contents lists available at ScienceDirect Expert Systems With Applications journal [homepage: www.elsevier.com/locate/eswa](homepage: www.elsevier.com/locate/eswa) | Generalizable deepfake detection framework using hybrid convolution-based EfficientNetB7 with attention mechanism | | | | --- | --- | --- | | Tirupathi Rajesh a,* , S. Maruthupermalb
a Department of Computer Science and Engineering, Bharath Institute of Higher Education and Research Selaiyur, Chennai 600127, India b Department of CSE, Bharath Institute of Higher Education and Research Selaiyur, Chennai 600127, India | | | | A R T I C L E I N F O
Keywords:
Deepfake detection
Video frames
Facial landmark extraction
Hybrid convolution network
EfficientNetB7 with attention mechanism Updated random parameter-based Fennec fox optimization | A B S T R A C T
The highly realistic fake videos have been created by the deepfake technology in recent years. The political and societal consequences occurred because of the misinformation disseminated by the fake videos. To prevent these issues, automated approaches for deep fake detection are necessary. The deep fake video is detected in this work using hybrid deep learning. The main contribution of the proposed research is to determine deepfakes from videos, which aims to avoid the spread of malicious rumours. The differences between the real and deepfake videos are effectively identified through this developed approach, as they utilize the different features from the facial landmarks along with the eye blinks for determining the sophisticated fakes from the authenticated visual contents. The required videos are collected from the public databases. The frames are extracted from the collected videos. The attained video frames are given to the facial landmark detection process, where the coordinates of the eyes, lips, and nose are extracted. From these detected facial landmarks, the texture features, shape features, and the number of eye blinks are obtained for better detection purposes. The extracted features are integrated with the optimized weights for attaining the weighted fused feature, where the Updated Random Parameter-aided Fennec Fox Optimization (URP-FFO) is used for tuning the weights optimally. Each frame of 1-dimensional data forms 2-dimensional original video frames, which are taken as feature set 1 from the weighted fused features. The 3-dimensional data from the facial images is considered as feature set 2. The attained features are given to the Hybrid 2D-3D Convolution-based EfficientNetB7 with Attention Mechanism (HC-EB7AM) for differentiating the real and fake videos. The manipulated contents are easily identified by this approach. Finally, several measures are used for validating the performance. The potential strength of the proposed model is validated by the experimentation with the accuracy of 93.66% using dataset 1 and 94.38% using dataset 2, respectively. | | 1. Introduction The emergence of inexpensive devices like cameras, smartphones, and computers is the reason for the development of social media (Alnaim et al., 2023). The exponential growth of social media helps to share content very quickly, which provides easy access to the shared data (Tan et al., 2022). The enormous progress in machine learning and deep learning algorithms is used to detect manipulated video content for preventing the spreading of misinformation, which protects people from danger (Khormali & Yuan, 2024). The spreading of deep fake images in the modern world threatens society very seriously. Deepfake is represented as synthesized video or audio content generated via artificial intelligence (Park et al., 2024). The categories of deepfakes include swap-face, facial feature manipulation, and face synthesis (Guo et al., Abbreviations: URP-FFO, Updated Random Parameter-aided Fennec Fox Optimization; HC-EB7AM, Hybrid 2D-3D Convolution-based Eff

FAQs

How does the method detect deepfakes from video content?

It extracts video frames and derives facial landmark coordinates (eyes, lips, nose) plus eye-blink information. These features are fused using optimized weights and fed into a hybrid 2D-3D EfficientNetB7 with attention to classify real versus fake videos.

What role does the Updated Random Parameter-based Fennec Fox Optimization (URP-FFO) play?

URP-FFO tunes the fusion weights for combining extracted facial texture, shape, and eye-blink features into a weighted fused representation that improves detection.

What detection performance is reported for the proposed model?

The model achieves 93.66% accuracy on dataset 1 and 94.38% accuracy on dataset 2 during validation using multiple performance measures.

Jul 30, 2026

Upload by Jiven