Deep Learning Approach for Detecting Fake Images Using Transfer Learning

The increasing sophistication of image manipulation tools poses a significant threat to the credibility of digital visual content. This study presents a deep learning-based classification model for detecting fake images by leveraging transfer learning using the ResNet50 architecture. The model architecture was fine-tuned to perform binary classification between real and synthetically generated images using a balanced dataset comprising 140,002 labeled images. A comprehensive pipeline was designed, including preprocessing steps such as image resizing to 224×224, normalization, augmentation (rotation, flipping, shearing), and validation splitting. The architecture incorporates layers that stabilize training and reduce overfitting while retaining robust feature extraction from the pre-trained base. Initial training yielded 80.85% accuracy and 0.78 F1-score, while post-fine-tuning the model improved to 81.35% accuracy and 88.39% Area Under the Receiver Operating Characteristic Curve, AUC. Empirical evaluations highlight the model's capacity to discern imperceptible manipulation patterns, enabling practical deployment in media forensics and secure image validation systems. Future enhancements may include attention-based classification, adversarial robustness, and region-specific forgery localization.