Sibos 2016: Cyber Attacks and AI Prediction

I kicked off the prestigious Sibos event by attending a conference given by Kalyan Veeramachaneni, Principal Research Scientist at the Laboratory for Information and Decision Systems, MIT.

The presentation was kicked off with Kaylan informing us that predictive analytics are both a necessity for multiple industries, but also that the nature of the work is very complex. Businesses and especially analysts will note that past occurrences within an attack will never happen again once the victim has become aware of it. The awareness of malicious attack patterns therefore means that the attackers will alter their behaviour, constantly. This nature has been coined as “non stationary date/patterns”.

During his speech, Kaylan, went over how the collaboration between MIT and PatternEx has resulted in the creation of a machine learning system which can successfully predict or find 86% of cyber attacks. By outlining how previous systems worked it paved a way for Kaylan to demonstrate just how much potential AI predictors could be. The old, most commonly used system is called the unsupervised learning system. The way this works is with a lot of manual work. The predominant issue with this system is the vast amount of outliers, potential threats, and the system cannot always be accurate as only analysts are trained to really identify whether or not a threat really is a threat. I turn this creates what is known as ‘alarm fatigue’ which is when the analysts become oo exhausted from checking out an infinite amount of potential threats.

The challenges for manual systems are many; a business has to find the right level of human input, have to receive expert sources, not generic crowdfunding sources, they have a limited bandwidth then have to ask themselves questions such as What information should we show? And How do we capture the most input?

 

By using the created virtual analyst alongside the outlier system the results can be improved drastically. Over the test period, three months, 86% of 315 attacks were predicted correctly and the goal for this technology is to create a way so the analysts can be shown as few events as possible, to avoid fatigue. Following the success of the tests PatternEx are now in the market to sell the technology and have already started selling it to certain banks.

 

 

Author: Dylan Jones

Share This Post On
468 ad