1. Face forgery detection via optimum deep convolution activation feature selection algorithm using expert-generated images.
- Author
-
Murtaza, Ghulam, Memon, Raheel Ahmed, Ali, Safdar, Shagari, Nura Modi, and Ali, Javed
- Subjects
FEATURE selection ,FORGERY ,CONVOLUTIONAL neural networks ,FUSIFORM gyrus ,ALGORITHMS - Abstract
The recent face forgery detection models have shown better accuracy for in-house generated and altered face image datasets. Therefore, the results of such a model cannot be compared directly and fairly. Moreover, some of the studies used datasets of smaller size, thus models can be overfitted. To address these issues, we selected a larger public dataset, created and altered by experts in photoshop. Thus, the goal of this study is to develop a face forgery detection model for expert-generated standard images. As a preprocessing step in our proposed approach, a convexhull reduced of an aligned face is extracted from each image. A convolutional neural network-based face forgery detection model is developed and trained from scratch. The images are normalized and label smoothing is applied to train the model to get minimum validation loss. The trained model is used to extract 1024 features of each image. Afterward, a feature selection algorithm is developed to select the optimum number of features without compromising the model performance. Moreover, random parameter optimization is performed to get the best parameters to train multiple classifiers. Only 36 features out of 1024 are selected using kNN(k = 4) for 76.04% sensitivity, 71.76% AUROC. The proposed approach achieved comparable results compared to the existing state-of-the-art face forgery models. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF