In the AI world, fairness is a paradox

If we as human beings are to accept artificial intelligence (AI) as part of our lives, then we have a few basic demands. First, we want AI systems to make fair decisions, free of any inherent bias. Second, we want AI systems to respect our privacy. Nokia Bell Labs holds these principles of fairness and privacy in the highest regard, to the point that both are included in our 6 Pillars of Responsible AI.

There’s only one problem.

Fairness and privacy are often opposing concepts in the AI world. To prevent bias, AI systems gather personal data on the people they are making decisions about, and this means letting go of privacy. On the other hand, once the privacy bubble is breached, that sensitive data may end up inadvertently reinforcing the same historical biases AI was meant to eliminate. This is the paradox between fairness and privacy.

Take the example of insurance. Like in many other industries, the practice of setting insurance premiums historically has been riddled with biases based on race, gender, age, sexual orientation, religion and socioeconomic class. An AI system intended to correct those biases, however, would want access to those very same attributes to ensure they aren’t used as the basis for discrimination. These attributes, though, are highly sensitive information and commonly not available in datasets. And if they are, we arrive at the paradox. As discrimination historically has been a key factor in driving decisions on where to set premiums, the AI begins basing its decisions on the same flawed criteria. Therefore, the AI system could perpetuate – or worse, amplify – the same historical injustices surrounding insurance access it was intended to correct.

What if we were to strip all personal data from the process as many regulators are now demanding? It stands to reason that if an AI system had no knowledge of race or gender, it couldn’t infer anything about race or gender, right? Well, that’s not how mathematics works. As this video from Minute Physics and the US Census Bureau wonderfully illustrates, AI systems can inadvertently determine people’s identities by mathematically matching individual attributes from different datasets. With that inferred personal data in play, AI systems again base their decisions on historically flawed criteria. But in these cases, we face an additional problem. As the AI system initially had no knowledge of race or gender, it’s impossible to know whether race and gender were factors in its decision making.

This raises a host of questions to how the AI world should handle privacy versus fairness, not the least of which is whether a technical solution that maintains privacy while promoting fairness is even possible. This challenge has become a major focus of Nokia Bell Labs AI research as we ultimately believe the AI field will never reach its full potential unless it fully adopts such ethical principles. Our current research builds upon what is now considered the most promising technique to privacy-preserving AI: differential privacy. By mathematically jiggering individual values for certain data (for instance altering the age of every person within a certain range), differential privacy is able to deliver an accurate portrayal of the larger dataset while obscuring the private data of individuals.

But this is just a starting point. While differential privacy does protect people’s personal data it doesn’t guarantee that all an AI’s decisions will be fair. You can think of differential privacy as a pair of thick goggles that obscures the fine details of personal data from the AI when it makes its calculations. The problem is the blurriness of the goggles goes in both directions. The AI system can’t tell if it has inadvertently discriminated over a particular group of individuals because it has no certainty over their personal data. Moreover, this blurriness can have a disproportionate impact over minority groups: the smaller the group of individuals the less accurate predictions the AI system will give.

To solve this problem, Nokia Bell Labs researchers have developed a means for AI systems to perform a form of meta-analysis on their own differential-privacy-based decisions. Using mathematical models, these techniques give more equal anonymity and provide fairer treatment for all individuals, regardless of whether some groups are underrepresented in its data – all without knowing what the defining characteristics of those groups are. For instance, when training an AI system, we can adjust how much personal information passes through the privacy goggles for each individual, balancing it against potential biases. We can perform this balancing on an algorithmic basis, or we can compute individual privacy parameters that allow more data to be collected from groups with weaker privacy protection. The result is an AI system that can minimize the trade-offs between fairness and privacy. We call the underlying technique that enables this balancing individual privacy accounting.

We are already introducing new differential privacy techniques into many AI systems and solutions being developed at Nokia Bell Labs. For instance, we are applying differential privacy techniques to deep neural networks, which, due to their scale, are particularly prone to reveal personal information from data sets. By carefully modifying the algorithms used to train neural networks, it is possible to obtain strict privacy guarantees that meet emerging regulations. We are also combining differential privacy with federated learning, which would allow each client in a decentralized collaborative-learning system to keep its local data secret from the other clients. As we perfect our differential privacy techniques, we plan to implement them in other AI systems as well, and eventually we hope to make these privacy-preserving technologies core features of Nokia products.

We want AI to make fair, unbiased decisions. At the same time, we want AI to maintain our digital privacy. These may seem like simple requests, but in mathematical reality, achieving both is an extremely complex problem. It’s a problem, however, our industry ultimately must solve. To create truly fair AI, these systems need some access to personal data. But that doesn’t mean all notions of privacy are thrown in the rubbish bin. At Nokia Bell Labs, we are creating the responsible AI systems that can use personal data to make just decisions, while still taking every measure to protect people’s privacy.