A Hybrid Deep Learning Approach: Conditional GAN and ResNet-18 for Counterfeit Rupiah Classification

The increasing circulation of counterfeit banknotes in Indonesia, stimulated by the rapid advancement of affordable digital printing technologies, threatens the credibility of the Rupiah and poses significant economic risks. Current detection practices mainly take manual procedures, relying on visual inspection or basic tools like UV lamps, which are insufficient for detecting sophisticated or partially altered counterfeit banknotes. To overcome these limitations, this study proposes a hybrid deep learning model that integrates Conditional Generative Adversarial Network (CGAN) and ResNet-18 using full-image of banknotes as inputs. CGAN facilitates the generation of labeled synthetic counterfeit data used in the training process, while ResNet-18 enhances the feature extraction capability of the model as a discriminator. A grid search was conducted to optimize hyperparameters, where the best-performing model achieved 99.33% accuracy on testing sets. Evaluation results demonstrate that the proposed model outperforms the vanilla CGAN with 81.33% accuracy and surpasses comparable studies in the field. This research underscores the potential of combining conditional generation with deep feature learning to build a robust, scalable, and automated counterfeit detection system for Rupiah.