IBM Wants to Enable AI Model Sharing and Data Privacy

IBM is developing data privacy and security solutions to defend against the misuse or unauthorized use of data. The company is building these technologies in its research labs in Silicon Valley, trying to leverage subsets of artificial intelligence, such as natural language processing and machine learning.

Wendy Belluomini, director of AI and cognitive software at IBM, tells WatersTechnology that financial institutions and technology companies use large volumes of user and client data to build and train AI models and algorithms, but in many cases, they might fail to consider that those doing the programming are not authorized to view that data.

“Banks have all of this data about you and they want to use that to build these AI models,” Belluomini says from the sidelines of the Sibos event in London.

“They want to use that data in a way that is legitimate; they want to market to you in a way that is based on how you are behaving and what your needs may be; but they don’t want to expose that data to someone that doesn’t need to see it,” she says.

Across many industries, including the financial markets, gaining access to data should only be permitted on a need-to-know basis, Belluomini says. This is difficult to achieve, however, in cases where data scientists or AI engineers want to use client data to extract insights or train complex models.

In response to these difficulties, IBM is building a variety of data privacy and security solutions to reduce the risk of breaching global data protection rules and cross-border data sharing laws.

The company is designing and developing AI models that comply with regulation and can be shared between financial institutions or when outsourcing to fintechs, without sharing the data used to program the technology. As AI models are based on sets of rules, Belluoimi says, there is a set of techniques that can be used to isolate the data from the model.

One method is to encrypt the data used to train the AI technology. Belluoimi says the idea is that these federated models will enable firms to partner with other industry firms in strengthening their AI capabilities by either trading models, merging models or leveraging a fintech to bolster the models.

Additionally, encrypting the data used to train AI technology will prevent internal teams from viewing the data unless permitted to with a key. Belluomini says data taken from various parts of the institution can be stored in a fully homomorphically encrypted database, which can be used to feed algorithms and intelligent technologies.

Additionally, IBM is also building an AI product that will monitor and protect against security breaches and cyberattacks.

“There are these issues around data being used in ways that people are uncomfortable with, but then again there are issues with data being stolen, data breaches, and we hear about these every few weeks,” Belluomini says. “So another application is to use AI to protect from those data breaches and defend your systems instead of depending on large numbers of security analysts trying to figure it out by hand.”

As an example: the surveillance technology can use NLP and knowledge graphs to identify keywords used in discussion forums, in communication networks and even in parts of the dark web. The AI will then alert security analysts of instances of potential misuse of data or cyber attacks.