An Explainable Counterfeit and Genuine Ethiopian Banknote Classification using Deep Learning

Abstract

Counterfeiting banknotes is a widespread and constantly evolving problem that presents substantial difficulties for financial institutions and law enforcement organizations on a global scale. The current approaches for counterfeit detection predominantly depend on the manual examination of security features and the utilization of hardware-based solutions, such as bill counters and detecting machines. Nevertheless, the cost of hardware-based solutions is capital expensive, while manual inspection is a laborious and subjective procedure that is susceptible to inaccuracies. The utilization of manual inspection or machine-based procedures is proving insufficient in addressing the challenges posed by constantly emerging threats. This study introduces an innovative approach for distinguishing between counterfeit and authentic Ethiopian banknotes. The study utilized advanced deep learning approaches, notably CNN, transfer learning and similarity learning. In the transfer learning approach, we used explainable AI (AI) techniques on our InceptionV3, VGG16, VGG19, and DenseNet121 models. We employed two explainable AI frameworks, namely Shapley Additive Explanations (SHAP) and TensorFlow Explain (TF EXPLAIN) for better understanding of the decision behind our classification models. With this approach, all the models had an accuracy beyond 99%. Notably, the Dense121 model exhibited a marginally superior accuracy of 100% and the proposed CNN model achieved relatively a very high accuracy of 98%. In the similarity learning approach, we incorporated similarity learning in conjunction with transfer learning to enhance the efficacy of our classification models. In this approach, we employed pre-trained models, namely Inception and VGG16, to construct our similarity models. Both models had a higher classification accuracy of 99.25% but the Inception model scored a better distance score of 0.335824. The experimental result suggests that exploring the use of AI as a novel approach to combat counterfeit Ethiopian banknotes is a worthwhile endeavor