Identification of Fake Indian Currency Using Convolutional Neural Network

Abstract

The circulation of counterfeit Indian currency continues to challenge the financial integrity and economic stability of the country, particularly in cash-driven transactions. Manual inspection and traditional detection systems often fail under real-world conditions due to limitations in generalization, speed, and scalability. This study aims to design and implement a real-time, lightweight counterfeit detection system that combines deep feature extraction using MobileNet with robust classification via Support Vector Machine (SVM) and Random Forest (RF) classifiers. The dataset includes 248 genuine and fake Indian currency images, which were augmented and preprocessed to improve generalizability. Feature extraction was performed using MobileNet’s depthwise separable convolutions, followed by classification using machine learning models trained with cross-entropy loss. The hybrid architecture was deployed using a local web server (XAMPP and Flask), making it accessible for real-time testing. Experimental results demonstrate that the MobileNet + SVM model achieved the highest performance, with 97.2% accuracy, an F1-score of 0.97, and a mean inference time of 1.8 ms per image, outperforming conventional CNNs and baseline machine learning methods. The model maintained over 90% accuracy under challenging image conditions such as blur, low light, and rotation. This work contributes a deployable, scalable solution to counterfeit detection and opens future avenues for expansion into multi-currency support, vision transformers, and adversarial robustness. It presents a significant step toward intelligent financial authentication systems for real-world applications