Abstract
At present, ATMs (Automated Teller Machines) are one of the essential
services for our daily life. It is also true that the thefts of false transactions and pin
thefts are increasing yearly. A significant amount of theft at ATMs is due to pin
overlooking and card skimming. Biometrics provide promising security but have high
implementation costs. Also, Indian laws discourage using BiometricsBiometrics in all
places. So, can AI be the solution to this problem? Instead of using keypad-based
inputs for pins, gesture detection with AI can be used for secure inputs. A trained deep
neural network can detect count from the hand symbols/gestures. The gesture input is
given by inserting the hand inside a safe box with a high-resolution camera attached.
The camera takes images and sends them to Raspberry Pi or any other embedded
system. The Raspberry Pi executes the lightweight ML model to detect the count. The
detected count is then encrypted and passed to the ATM. Using a gesture identification
system removes the problem of pin theft and can be developed and implemented with
the slightest modification in ATMs. In the current COVID period, execution of ATM
works with minimum contact to public surfaces has increased immensely. In this
system, a keypad is also removed and can further be incorporated to read a variety of
inputs from gestures instead of just hands. This chapter explores how lightweight
neural networks can be trained to detect sensors and run on low-processing systems
like Raspberry Pi. We achieved an accuracy of 94%-97% in detecting gestures and pins
where accuracy varies for each motion.
Keywords: ATM Security, CNN, Computer Vision, Deep Learning with Embedded System, Gesture Detection, Mobile Net, Raspberry Pi, Transfer Learning, Tensor Flow, VGG model.