Fake Currency Detection Using Machine Learning for Accurate and Reliable Banknote Authentication System

Abstract

The increasing accessibility of high-resolution colour printing has intensified the problem of counterfeit banknotes. This study presents a deep-learning framework for classifying Indian currency images as real or counterfeit using the Xception architecture. The reported dataset contains 232 images of genuine and counterfeit notes and is described primarily in relation to the INR 500 and INR 2,000 series. Images were resized to 224 x 224 pixels, normalized, augmented, and divided into training, validation, and testing sets. Xception was initialized with ImageNet weights and fine-tuned, while a Python and Flask interface was used to accept images and display the predicted class. Performance was assessed with accuracy, precision, recall, F-measure, a confusion matrix, an accuracy curve, and functional test cases. The thesis reports a training accuracy of 93.34% and a validation accuracy of 97.00%. The deployed performance page reports accuracy of 0.97, precision of 0.94, recall of 0.97, and F-measure of 0.97. The confusion matrix records 56 counterfeit notes correctly classified, three counterfeit notes classified as real, no real notes classified as counterfeit, and 48 real notes correctly classified. The results indicate that Xception provides an effective basis for time-efficient counterfeit-currency detection, while further work is required to broaden denomination coverage, incorporate security features, and improve robustness to noisy images.