CS17 - An Efficient Deep Learning Approach for Deepfake Detection
Abstract
The rapid advancement of generative deep learning models has triggered a surge in highly realistic deepfake images and videos, threatening information integrity, digital security, personal privacy, and public trust. With synthetic content now nearly indistinguishable from the real thing, the need for robust and automated detection systems has become more critical. This study presents a dual-branch deepfake detection framework that combines an EfficientNet-B4 spatial branch with a convolutional neural network (CNN)-based frequency branch, evaluated on the FaceForensics++ (FF++) dataset. Frame-level spatial features were extracted using EfficientNet-B4, while frequency-domain features were derived from Fourier-transformed images using a dedicated CNN branch. These complementary representations were fused through weighted feature concatenation and classified using a multilayer perceptron. Data augmentation was applied to improve generalization and robustness across manipulated samples. When evaluated on the FF++ dataset, the proposed model achieved an accuracy of 99.38%, an AUC of 99.96%, a precision of 99.57%, a recall of 99.19%, and an F1-score of 99.38%. These results demonstrate that the proposed approach offers a reliable and computationally efficient performance for deepfake detection, highlighting its potential for practical deployment in safeguarding the integrity of digital media.
CS17 - An Efficient Deep Learning Approach for Deepfake Detection
University Readiness Center Greatroom
The rapid advancement of generative deep learning models has triggered a surge in highly realistic deepfake images and videos, threatening information integrity, digital security, personal privacy, and public trust. With synthetic content now nearly indistinguishable from the real thing, the need for robust and automated detection systems has become more critical. This study presents a dual-branch deepfake detection framework that combines an EfficientNet-B4 spatial branch with a convolutional neural network (CNN)-based frequency branch, evaluated on the FaceForensics++ (FF++) dataset. Frame-level spatial features were extracted using EfficientNet-B4, while frequency-domain features were derived from Fourier-transformed images using a dedicated CNN branch. These complementary representations were fused through weighted feature concatenation and classified using a multilayer perceptron. Data augmentation was applied to improve generalization and robustness across manipulated samples. When evaluated on the FF++ dataset, the proposed model achieved an accuracy of 99.38%, an AUC of 99.96%, a precision of 99.57%, a recall of 99.19%, and an F1-score of 99.38%. These results demonstrate that the proposed approach offers a reliable and computationally efficient performance for deepfake detection, highlighting its potential for practical deployment in safeguarding the integrity of digital media.