How machine learning can help with demonetisation and counterfeit currency

It has been more than a year since the Prime Minister of India announced to the nation, the demonetisation of certain Indian currency that accounted for nearly 86 percent of the cash in circulation. Much has been written about it, including the infamous tweet by Shashi Tharoor likening the global domination of the Indian currency to the recently crowned Miss World.

Machine learning can help in combating counterfeit currency. Reuters.

Perhaps the time is ripe to take a step back from the financial and social ramifications of this bold step by Narendra Modi, and take a moment to look at certain technical aspects instead. How do we measure the impact of demonetisation or prevent the system from being re-polluted with counterfeit currency notes?

Machine learning to the rescue.

The use of machine learning has been greatly touted in multiple applications in the banking and financial services industry. Its use in demonetisation and detection of counterfeit currency is yet another interesting avenue to explore.

At first glance, it does not appear that machine learning could be used in a significant way to help with demonetisation. However, consider the vast amount of data that was collected in terms of the cash deposits made in various banks and accounts across the entire country. If we could collate all this rich data, various interesting patterns and actionable insights will arise.

For example, we can track which regions showed a large number of cash deposits upon demonetisation. Further drilling down into bank branches and account owners, we could run some collaborative filtering algorithms to cluster together similar users and identify compelling transactional patterns. In the future, these patterns could help prevent potentially suspicious parallel economy transactions.

Another application of machine learning that we have already seen is the evaluation of public feedback via online media regarding demonetisation. Sentiment analysis using natural language processing techniques have enabled researchers to evaluate about more than 6,000 tweets made on the day of the announcement. Results showed that about 65 percent of people welcomed this change with the largest supporters being in the age group of 24 to 34 years.

Clearly, the change was predominantly popular with the millennials, although we must take these results with a pinch of salt since these are the very people who are more avid users of social media.

Turning to counterfeit currency, a study conducted by the Indian Statistical Institute (ISI) Kolkata in 2016, revealed that 250 out of every 10 lakh notes in India are counterfeit. At any given point of time about 400 crores worth of counterfeit notes are in circulation, and another 70 crores are infused in the system every year.

Law enforcement agencies are able to intercept only about one-third of these fake notes. A cost-effective and viable solution to detect these fake notes is the need of the hour to ensure that we do not nullify the advantages that demonetization introduced in the national financial system.

Currency notes are already equipped with anti-counterfeiting characteristics; to name a few, watermarks which are faint designs visible only when held against light, optically-variable ink which displays two distinct colors depending on the viewing angle, intaglio printing where the surface of the note in sunken to hold the ink, see-through features which are a combination of two independent background and foreground images, and latent images which are again visible only at a certain angle.

Any solution that can be used by an end-user for detection of fake currency should have certain desirable features: no damage should be caused to the underlying note, a mobile-based solution would be most cost-effective and provide real-time usability, and microscopic material detection and advanced image recognition with micro-level precision would ensure near-perfect accuracy.

Photographs (which do not damage the notes) can be easily taken using ubiquitous smartphones equipped with high-resolution cameras. However, a smartphone camera (which is often hampered by poor illumination) may not suffice to capture the details of anti-counterfeiting features. Provided we can capture these details, we then need to develop a machine learning solution that can evaluate these high-resolution photographs using a sophisticated image recognition system.

Image recognition has grown by leaps and bounds in the past couple of years by training deep neural networks to classify images into thousands of different categories. These pre-trained networks such as AlexNet and GoogLeNet can be used to recognize a new set of images using a smaller data set, and the use of transfer learning where the knowledge obtained from one domain of images can be transferred to classify new images from another domain.

We can train a deep learning model to recognize real and fake currency notes using a limited set of both types of note images. This model can be easily ported on a smartphone using light-weight software packages such as TensorFlow Mobile. This low-cost option, which does not need an internet connection upon deployment, makes it attractive to develop.

As we ponder over the premature concerns about threats posed by artificial intelligence to humanity at large, we have presented a concrete example where we can use it for the betterment of society. The primary hurdles to the development of such a solution is the availability of fake currency notes for training the machine learning system, and developing lightweight, low-cost hardware to capture high-resolution photographs.

With due help from the regulatory authorities, it should be an easy one to overcome the former hurdle. I recommend that banks and governmental financial agencies invest in the research and development of such a solution to sustain the transient advantages obtained from demonetization on a long-term basis.