Counterfeit Drug Detection Using Deep Learning and Synthetic Data Generation: A Hybrid Approach

Abstract

Counterfeit pharmaceuticals pose a critical threat to global public health, particularly in low- and middle-income countries where regulatory oversight and access to advanced authentication tools are limited. This study presents a comprehensive deep learning framework for automated counterfeit drug detection that addresses the persistent challenge of limited authentic counterfeit training data through synthetic data generation. A hybrid dataset was constructed by combining real-world pharmaceutical images from public repositories with synthetically generated counterfeit images produced using Stable Diffusion XL. Five convolutional neural network architectures—ResNet50, InceptionV3, VGG16, EfficientNetB0, and MobileNetV2—were evaluated under real-only, synthetic-only, and hybrid training regimes. The hybrid approach achieved the best performance, with ResNet50 attaining 97.0% accuracy and an AUC-ROC of 0.993, significantly outperforming models trained on real data alone. Integration of optical character recognition for batch number and expiration date verification further improved system accuracy to 98.2%. A mobile application prototype was developed to demonstrate deployment feasibility, achieving sub-200 ms inference time after model optimization. The results confirm that carefully designed synthetic data generation, when combined with real-world data, substantially improves robustness, generalization, and practical usability of AI-based pharmaceutical authentication systems.