Citi Examines Machine-Learning Risks


New technologies bring new risks, and machine learning is not the exception, according to experts.

The technology has seen a dramatic uptick in interest since cloud-based computing and storage driven the price of computing to a historic low.

Over the past decade the number of machine-learning research papers has grown by 7,300 percent compared to the 1,300% growth in research regarding general artificial intelligence topics, said Yogesh Mudgal, director emerging technology risk at Citi, during the AI Summit held in Manhattan.

“The number of research papers is not important, but what it shows is that the ecosystem around machine learning is robust, strong, and growing,” he said.

However, Mudgal noted that there is a significant gap between implementing the technology and managing its associated cybersecurity risk.

“Machine-learning implementations are still very nascent compared to traditional technologies,” said Mudgal. “I believe, as a community, that this is the best time to develop a strategy for better and sounder communication.”

For its part, Citi is engaging the various organizations and stakeholders within the bank who have or plan to implement machine learning to develop a risk-management framework. The bank has gone as far to release a paper on the topic to spur conversation between firms.

The conversation’s goal is not to throttle innovation but to enable innovation via a set of yet to be developed industry best practices.

Firms considering deploying machine-learning instances should determine if they need to use the technology to accomplish their goals, according to Mudgal. Once they have, they should establish a living document to manage the risk throughout the project’s lifecycle.

“From creation, testing, and deployment to decommissioning, everything should have some report that says what the organization is doing at every step,” he added.

The document should contain what the firm’s plans on what to do with the data, such as purge it or add it to the company’s inventory, when it decommissions an instance of machine learning.

More importantly, organizations need to answer is who owns machine learning, according to Mudgal. “Is it the business owner, application manager, the developer, or the person providing the training data?”

Once companies address these foundational issues, they can begin a conscious effort examine machine learning’s limits from a risk perspective, he added.

Related articles