Deep Learning Based Fake Stamp Detection

Philately or Stamp Collection also known as the King of Hobbies. Philatelic fakes and forgeries are labels that look like postage stamps but have been produced to deceive or defraud. Learning to identify these can be a challenging branch of philately. Those who produce counterfeits appeal to a very different market from philatelists. They depend on their stamps being produced in large quantities in order to be able to recover their outlay. Political and propaganda forgery is produced by countries in conflict to hurt the opponent. Stamps may be issued to deprive the enemy of revenue, to distribute propaganda material, to cause confusion, and to depict propaganda messages. Sale of fake stamp papers mean a loss of revenue to the tune of crores of rupees to the exchequer. Stamp duty is an important source of revenue for the government. The state governments may have to probe the large financial transactions over the past decade to ascertain actual damage. Further, the negative effects on the financial sector would be manifold. It is very difficult to identify a fake stamp from a real one. It takes years of experience and study to be able to become an expert in philately and to be able to identify fake stamps. Postal services developed, early on, measures to protect the integrity of their stamps. Some of these steps are similar to those used to protect against forged currency. Major steps include Watermarks, Special paper, Delicate engraving, Printing methods, Special ink for postmarks, Insertion of silk threads, Secret marks either visible or invisible to the microscope, and Re-issue of stamps. It may not be possible to distinguish between philatelic and postal forgery if the stamps are unused, merely by looking at them; the techniques used in producing them are identical. Hence this paper aims on identifying whether the input given is a fake stamp. In deep learning, a Convolutional Neural Network (CNN) is a class of Artificial Neural Network (ANN), most commonly applied to analyse visual imagery. When an input is given, the deep learning model created will identify whether the input is an image of a stamp or something else. If the input given is identified as image of stamps, then it identifies whether the stamp is a real or a fake. If the stamp is real the model correctly identifies the country to which it belongs by reading the name of the country in the stamp. Stamps belonging to six countries over a period of 50 years were collected from the internet, and trained through the deep learning model. As there is no dataset available for fake stamps, they were generated by modifying the real stamps. The designed deep learning model is simple with accuracy of 99% would be very helpful in eliminating the fake stamp.