Auguring Fake Face Images Using Dual Input Convolution Neural Network

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology, utilizing auto-encoders and generative adversarial networks, poses challenges in identifying manipulated media.
• Existing methods struggle with detecting deepfakes under compression, blurring, or scaling, creating a significant research gap.
• The proliferation of deepfakes raises concerns regarding fraudulent activities and security vulnerabilities.

Purpose of the Study:

• To develop a robust deepfake detection model capable of accurately identifying counterfeit faces.
• To address the limitations of current detection techniques in handling degraded image quality.
• To enhance the reliability of digital forensics and security by improving deepfake identification.

Main Methods:

• Proposed a novel dual input convolutional neural network (DICNN) model.
• Employed ten-fold cross-validation for rigorous model evaluation.
• Integrated SHapley Additive exPlanations (SHAP) for explainable AI (XAI) to interpret model decisions.

Main Results:

• Achieved high average accuracies: 99.36% (training), 99.08% (testing), and 99.30% (validation) with minimal standard deviation.
• Demonstrated the model's effectiveness through SHAP analysis, providing visual explanations for its predictions.
• The DICNN model significantly outperformed existing state-of-the-art methods in deepfake detection.

Conclusions:

• The proposed DICNN model offers a highly accurate and reliable solution for deepfake image detection.
• The integration of XAI enhances the model's transparency and trustworthiness for forensic and security applications.
• This advancement is crucial for combating the misuse of deepfake technology and bolstering digital security.