Abstract:
Counterfeit currency is a major threat to the economic integrity and security of Bangladesh,
particularly with regard to high-denomination notes, including 500 and 1000 taka bills. The
current study sought to achieve an understanding of the modern deep learning approach’s
potential to effectively identify counterfeit currency. Therefore, we tested four widely
recognized Transfer Learning models available with pre-trained weights, including VGG16,
Xception, ResNet50, and DenseNet201. We trained these models on a large dataset of authentic
and fake images of Bangladeshi banknotes and assessed their capabilities to detect whether
banknotes are authentic or counterfeit. The DenseNet201 model demonstrated the greatest
identification power and accuracy according to the results, with an accuracy level of 97.69
percent. From the other models, Xception demonstrated 94.77 percent accuracy, VGG16
demonstrated 94.26 percent accuracy, and ResNet50 demonstrated 92.03 percent accuracy.
Regardless, the outstanding efficiency of the DenseNet201 model shows that it can be used as
a powerful tool for combating counterfeit in the country, providing huge strides over existing
technologies. Specifically, indeed, present study lends empirical evidence that deep learning
has a significant disruptive potential in the three-years future of financial security. It can
encourage the development of more advanced systems for detect counterfeit currency, which
can revolutionize the fight against financial crime and protect Bangladesh’s national economic
interests
