Fig 1.
Block diagram of the proposed ARC-Net framework integrating residual and attention modules for deepfake image detection.
Table 1.
Dataset split details for DeepFake datasets used in ARC-Net evaluation.
Fig 2.
Sample images of extended dataset:
(a) Real Images and (b) Fake Images.
Table 2.
Preprocessing and Data Augmentation Techniques with Parameters.
Table 3.
Model Parameter details.
Table 4.
Notation and Symbols Used in Algorithm 1.
Fig 3.
Overall architecture of proposed ARC-Net showing integration of residual and attention blocks atop EfficientNet-B0.
Fig 4.
Architecture of the Grad-CAM module used for visualizing attention regions with the ARC-Net model.
Table 5.
Result Evaluation of All Models.
Table 6.
Classification Report of Models.
Fig 5.
(a) Training/validation accuracy over 30 epochs, and (b) ROC curve showing high AUC score on the ‘140k Real and Fake Faces’ dataset.
Fig 6.
DenseNet121 performance analysis.
(a) Training/validation accuracy over 30 epochs, and (b) ROC curve on the ‘140k Real and Fake Faces’ dataset.
Fig 7.
Confusion matrices illustrating the classification performance of ARC-Net.
(a) and DenseNet121 (b) on the ‘140k Real and Fake Faces’ dataset.
Table 7.
Result Evaluation of ARC-Net model on Deepfake Dataset and Deepfake Database.
Fig 8.
(a) Train validation accuracy plot and (b) Receiver operating characteristic curve of ARC-Net on the Deepfake dataset.
Fig 9.
(a) Train validation accuracy plot and (b) Receiver operating characteristic curve of ARC-Net on the deepfake database.
Table 8.
Comparison of classification performance across various research studies on the DeepFake face detection with our proposed ARC-Net model.
Fig 10.
Confusion matrix of ARC-Net in (a) deepfake dataset and (b) deepfake database dataset.
Fig 11.
Correct and incorrect test examples from the deepfake dataset with ARC-Net.
Table 9.
Model Configurations and Experimental Setup.
Table 10.
Comparison of Model A and Model B on the Held-Out South Asian Test Set (100 Real Images).
Fig 12.
False Positive Rate change and McNemar contingency table between Model A and Model B.
Fig 13.
Confidence distributions of mean probability of real for Model A (blue) and Model B (orange).
Table 11.
Dataset Distribution and Results of OOD Evaluation 01.
Fig 14.
Performance comparison of OOD-01.
Table 12.
Dataset Distribution and Results of OOD Evaluation 02.
Fig 15.
Performance comparison of OOD-02.
Table 13.
Experiment Configurations.
Table 14.
Performance of ablation variants and the full ARC-Net model on the hybrid dataset.
Fig 16.
Confusion Matrix of ARC-Net analysis:
(a) Baseline EffecientNetB0 (Experiment A), (b) Attention bassed CNN (Experiment B) and (c) Residual CNN (Experiment C)in hybrid dataset. (d) DenseNet121-based CNN with residual and attention modules (Experiment D).
Fig 17.
Grad-CAM Explanation and Heatmap with ARC-Net.
Fig 18.
LIME Explanation for ARC-Net.
Fig 19.
Grad-CAM Explanation and Heatmap for DenseNet121, ResNet50, and ARC-Net on real human faces.
Table 15.
Representative IoU, Dice, and Peak-Activation Scores for a Subset of Test Images. The complete dataset comprises 200 images; only a concise selection is included here for clarity. Aggregate metrics (mean and 95% CI) are computed over the full test set.