Adversarial Attacks in Banknote Recognition Systems

Numerous computer vision algorithms have been designed to help detect the values of the paper banknotes for blind and visually impaired individuals. Previous algorithms relied mostly on traditional methods that are less efficient, accurate, and transferable than a convolutional neural network. So in our study, we used the convolutional neural networks to perform banknote image classification. We trained our network with 1,000 Thai banknote images. Our network could classify real-life banknotes used in our study with an accuracy of 100%; however, we have also found out that our convolutional neural network model is vulnerable to adversarial attacks. We then retrained the network using adversarial training which reduced the attack success rate by approximately 50%. Overall, because we did our experiments using MobileNet, which is a relatively small convolutional neural network that can be used in a mobile phone, we have trained an accurate and robust banknote recognition convolutional neural network that could be integrated into a mobile app or wearable device for the visually impaired.