An Optimized Deep Learning Pipeline for Image Authenticity Detection Using Multi-Stage Deep Learning Pipeline

DOI: https://doie.org/10.10399/JBSE.2026880368

Noorshaba Khatoon, Aditi Sharma, Divyanshu Chauhan

Keywords:

Deep Learning, Image Classification, AI-Generated Images, Real vs Fake Detection, Image Authenticity Detection, Synthetic Image Detection

Abstract:

The rapid development of generative AI has produced very convincing fake images, making it difficult to tell what's real or fake. This situation presents major challenges for digital forensics, detecting misinformation and verifying media authenticity. Because of this, we've developed a deep learning classification pipeline for distinguishing between natural and computer-generated images, using the DeepDetect dataset with over 1,00,000 images from a variety of categories. We used a transfer learning approach with the VGG16 back-end, and we finetuned the feature extraction through a series of multi-stage fine tunings. The Adam optimizer was used to optimize performance, utilizing multiple different learning rates and a binary cross entropy loss function incorporating label smoothing. Extensive hyperparameter tuning and data augmentation were performed. The model performance was evaluated with several evaluation metrics such as accuracy, precision, recall, F1- score, and confusion matrix. Our average overall accuracy was 99.13%. Our work shows how deep learning pipelines can reliably detect image authenticity.

PDF