DEEPFAKE IMAGE RECOGNITION VIA FEATURE LEARNING IN CONVOLUTIONAL NEURAL NETWORKS

Abstract

The surge in synthetically generated facial imagery— commonly termed deepfakes—poses escalating threats to information integrity and digital forensics. This paper presents a Convolutional Neural Network (CNN) framework employing transfer learning from the MobileNet backbone, pre-trained on ImageNet, for binary classification of real versus GANsynthesized facial images. A lightweight classification head comprising Global Average Pooling, a Dense(128) layer with ReLU activation, Dropout(0.5) regularization, and a Softmax output is fine-tuned on the Kaggle Deepfake-and-Real-Images benchmark dataset. Training for five epochs with the Adam optimizer (lr = 3×10⁻⁴) achieves 96.1% validation accuracy and a weighted F1-score of 0.960, surpassing several published baselines. An AUC-ROC of 0.991 confirms strong discriminative calibration. A Flask-based REST API and React.js frontend are deployed to enable real-time, browseraccessible inference.