Article: Detection of deepfake technology in images and videos Journal: International Journal of Ad Hoc and Ubiquitous Computing (IJAHUC) 2024 Vol.45 No.2 pp.135 - 148 Abstract: In response to the low accuracy, weak generalisation, and insufficient consideration of cross-dataset detection in deepfake images and videos, this article adopted the miniXception and long short-term memory (LSTM) combination model to analyse deepfake images and videos. First, the miniXception model was adopted as the backbone network to fully extract spatial features. Secondly, by using LSTM to extract temporal features between two frames, this paper introduces temporal and spatial attention mechanisms after the convolutional layer to better capture long-distance dependencies in the sequence and improve the detection accuracy of the model. Last, cross-dataset training and testing were conducted using the same database and transfer learning method. Focal loss was employed as the loss function in the training model stage to balance the samples and improve the generalisation of the model. The experimental results showed that the detection accuracy on the FaceSwap dataset reached 99.05%, which was 0.39% higher than the convolutional neural network-gated recurrent unit (CNN-GRU) and that the model parameter quantity only needed 10.01 MB, improving the generalisation ability and detection accuracy of the model. Inderscience Publishers - linking academia, business and industry through research

Title: Detection of deepfake technology in images and videos

Authors: Yong Liu; Tianning Sun; Zonghui Wang; Xu Zhao; Ruosi Cheng; Baolan Shi

Addresses: College of Cyberspace Security, PLA Strategic Support Force Information Engineering University, ZhengZhou 450001, Henan, China ' Research Institute of Intelligent Networks, Zhejiang Lab, Hangzhou, Zhejiang 311121, China ' College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, Zhejiang, China ' College of Cyberspace Security, PLA Strategic Support Force Information Engineering University, ZhengZhou 450001, Henan, China ' College of Cyberspace Security, PLA Strategic Support Force Information Engineering University, ZhengZhou 450001, Henan, China ' College of Engineering and Applied Science, University of Colorado Boulder, Boulder 80309, Colorado, USA

Abstract: In response to the low accuracy, weak generalisation, and insufficient consideration of cross-dataset detection in deepfake images and videos, this article adopted the miniXception and long short-term memory (LSTM) combination model to analyse deepfake images and videos. First, the miniXception model was adopted as the backbone network to fully extract spatial features. Secondly, by using LSTM to extract temporal features between two frames, this paper introduces temporal and spatial attention mechanisms after the convolutional layer to better capture long-distance dependencies in the sequence and improve the detection accuracy of the model. Last, cross-dataset training and testing were conducted using the same database and transfer learning method. Focal loss was employed as the loss function in the training model stage to balance the samples and improve the generalisation of the model. The experimental results showed that the detection accuracy on the FaceSwap dataset reached 99.05%, which was 0.39% higher than the convolutional neural network-gated recurrent unit (CNN-GRU) and that the model parameter quantity only needed 10.01 MB, improving the generalisation ability and detection accuracy of the model.

Keywords: deepfake technology; fake image and video detection; transfer learning; parameter quantity; detection across datasets.

DOI: 10.1504/IJAHUC.2024.136851

International Journal of Ad Hoc and Ubiquitous Computing, 2024 Vol.45 No.2, pp.135 - 148

Received: 31 Oct 2023
Accepted: 15 Dec 2023
Published online: 22 Feb 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Detection of deepfake technology in images and videos

Keep up-to-date