CNN EfficientNetB0 Approach Based on Transfer Learning for Deepfake Detection and Disinformation Mitigation

Abstract

Rapid advances in deep learning technology have led to the emergence of artificial intelligence (AI) media that is very similar to reality, called deepfakes, which have the potential to pose a serious threat to information integrity and public trust. Although detection methods using Convolutional Neural Networks (CNN) have been developed, most still struggle with generalization, particularly in distinguishing modern deepfakes from non-standard original images such as selfies, which often leads to high false positive rates. This study introduces a robust detection model based on the EfficientNetB0 architecture implemented through transfer learning techniques. To improve generalization capabilities and minimize bias, we compiled a large and balanced combined dataset by combining three different public datasets (including classic deepfakes, face swaps, and many authentic selfies). The model was trained using a two-stage strategy: first for feature extraction, then refinement with a very low learning rate. The model's performance was thoroughly evaluated on stratified test data using five key metrics. The results of the experiment showed outstanding performance, achieving 99.81% accuracy and a Macro F1 score of 99.81%. Additionally, the reliability metrics ROC-AUC, Average Precision (AP), and True Positive Rate (TPR) all reached 99.99%, while the False Positive Rate (FPR) remained strictly at 1%. As proof of concept, this optimized model was implemented in a web prototype built using the Django framework, allowing users to upload images and receive classification results in real-time.