Select Page

Last week we explored three examples of how Machine Learning (ML) is being used to innovate industry. This week we’ll highlight those same examples, and discuss how using the full AI toolbox augments the performance of an ML-only approach.

Revisiting ML in Use

Cyber Security: Nathaniel Gleicher, Illumio’s Head of Cybersecurity Strategy, recently presented at USENIX Enigma 2017.  (You can watch the full talk here.)  In the talk, Nathaniel discussed how, on average, an intruder can spend 146 days in a network without being detected.  In addition, most zero-day attacks happen at least 310 days prior to your organization becoming aware.  (Andy Greenberg (Forbes). “Hackers Exploit ‘Zero-Day’ Bugs for 10 Months On Average Before They’re Exposed.” October 2012.)  Nathaniel provides a compelling, real-world example in his talk, using the Secret Service to depict control within an environment. ML is a learning-based approach; as with all learning-based approaches, if you or I can learn something, what is precluding an adversary from learning as well?  Instead of learning the rules of a given domain, the Secret Service defines their environment and imposes their own set of rules to establish a safe infrastructure.  The same idea can be applied to cybersecurity, with other AI tools bolstering the security framework initially established through ML.  Graph Theory (GrT – using mathematics to model relationships between objects) can be applied to control specific attack paths, and Game Theory (GT) can formulate pre-emptive strategies for how an intruder might breach the defenders’ environment.  While ML still has relevance, only through combining an array of AI tools into one united defense can you truly counteract the presence of cyber criminals.

Graph Theory: Study of graphs (Diagram representing a system of connections among two or more things. Examples: Road Network, LinkedIn connections)

Banking: Banks are using ML to innovate their infrastructure and reduce internal costs.  According to Bloomberg, if current models remain valid, investment software – facetiously referred to as “robo-advisers” – will maintain $2 trillion in assets by the end of 2020.  (Regan, Michael P. Robo Advisers to Run $2 Trillion by 2020 if This Model Is Right. Bloomberg. Web. June 18 2015.)  While ML is adequate to make intelligent stock picks, this approach could be reinforced by incorporating elements of GT.  Specifically, if robo-advisers consider what their counterparts (e.g. other robo-advisers) will do in light of current information, they could exploit that intel, outstrip the competition, and lead the market.  This would allow robo-advisers to make decisions on how the market will behave, based on the ML models being learned by their counterparts.

Game theory: Study of how and why people make decisions. Chess masters calculate three to five moves ahead depending on the positions of the pieces.

Financial Services: A major challenge looming over the financial services industry is knowing how much capital to maintain at any given time to satisfactorily address developing situations.  If an X number of investors removed their funds at the same time, how much money would the firm need to pay out all claims while still optimizing solvency and the ability to generate interest on the untouched capital?  Firms might hold onto more funds than is required, and in the process, lose millions in ancillary revenue each year.  Financial firms use ML to understand trends in their Big Data, but ML is insufficient to understand more than ongoing trends and predictions sourced from those trends.  A more robust approach would incorporate using Bayesian Networks (BN – an AI tool offering the ability to probabilistically model random variables and their dependencies, i.e. determine the relationship between hair color and eye color) to reason over dependencies.  In conjunction with BNs, the use of Decision Theory (DT) would help understand both the glaringly obvious and hidden uncertainty in the market at any given inflection point.

Bayesian Networks: Directed graph with no cycles for modeling situations with uncertainty

This entry concludes Journey Into AI, our first blog miniseries.  Next week we’ll shift our focus to another pertinent topic and begin a new blog series on AI usage in the security domain.  While different in flavor, the content will remain just as accessible and intriguing.  We’re excited for you to come along on this journey.  

And if machines ever become sentient and turn against us, we like to think you’ll be better prepared for the future.