Deepfake Image Forensics Using Transfer Learning

Over the past few years, deepfakes have emerged as a serious threat to digital security, privacy, and information integrity due to rapid advances in generative models that produce highly realistic facial images. Manual verification has become increasingly unreliable, necessitating robust automated detection techniques. This paper presents a deep learning-based framework for distinguishing real and AI-generated face images using a transfer learning approach built on DenseNet-121. Rather than introducing a complex architecture, the proposed method emphasizes an effective two-stage, curriculum-based training strategy to enhance accuracy and robustness under realworld distortions. Experiments conducted on a large-scale dataset of approximately 140 K images demonstrate strong performance across standard classification metrics, achieving an accuracy of 96.8% and an ROC-AUC of 98.2%. The trained model is further integrated into a lightweight web-based application that provides real-time image authenticity prediction with confidence scores, highlighting its practical applicability.