Center of Applied AI
At Onfido our mission is to build an open world, where identity is the key to access. A world where anyone will be able to remotely verify their real identity with an identity document and their facial biometrics. Because the more people are able to remotely verify their real identity, the more accessible essential services become.
We believe that artificial intelligence is one of humanity's most powerful tools to solve world issues around identity and access. But it also has the capacity for harm. We are committed to ensuring we build fair AI in an ethical way.
The Center of Applied AI’s goal is to develop best practices, and build high-performance artificial intelligence and machine learning powered systems. Systems that power mission critical applications, and operate in the real world.
AI designed for reality
When most people think of ‘Artificial Intelligence’ they likely think of something which is both perfect and independent. A supreme power, which eclipses the human brain in every way. But in actuality, we’re a far cry from developing this kind of ‘General AI’.
Instead, we've been able to create more specific AIs, thanks to advances in technology, ever-increasing data availability, and algorithmic advances. These systems still outperform humans, but only when applied to specific tasks. These ‘Applied AIs’ are what allow us at Onfido to automate our identity verification services to a high degree of accuracy, operating faster than human experts alone ever could.
Building high-performance machine learning systems
Our machine learning systems are trained with real-world data, so they don’t just perform on paper. Our challenge is to make them as close to perfect as possible in imperfect conditions.
Our systems are built to mission-critical standards. This means that every mistake carries a high cost, meaning 99% performance is not good enough. They are built to operate to outlier measures. To achieve this we need to understand how they’re making decisions to continually optimize their performance. Our models are built with exception paths and innovative systems which mean they are more than ‘black boxes’, so we can continue to strive for the highest performance.
Want to learn more?
As our mission is to make identity the key to access, we must do so responsibly and ethically. That’s why Onfido follows a number of ethical practices when working with AI, to make sure it operates in an impartial and considered way.
Mitigating algorithmic bias
‘Algorithmic bias’ describes systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one group of users over others.
Many things can trigger bias: from the design of the algorithm itself to unintended or unanticipated ways of using it, or even the decisions about how we use data to train the algorithm.
We believe that algorithms need to operate globally and to the same standard for everyone.
We’re proud to be participating in the UK's first Information Commissioner's Office (ICO) data protection and privacy sandbox to systematically measure and mitigate algorithmic bias in machine learning models. In particular, we’re focusing on racial bias and other data-related bias effects for remote biometric-based identity verification. The latest results represent an overall 10x improvement from the previous algorithm version, and a 60x false acceptance improvement for faces of African descent.
Dealing with data
At Onfido we process two key types of data when performing an identity verification: identity documents and facial images. It’s critical we treat this data with the utmost care. We have codified the principles of how we handle data through our internal privacy program. We continually reassess our practices to make them as responsible and secure as they can be. For example, we operate on a zero-trust policy, and treat every access attempt as an attack. We also protect the personal data we have permission to use for research, containing it in a separate, monitored and highly secure data research environment, called SILO. We have enabled self-service deletion and rolling deletion of individual records for our clients, available directly from their dashboard and via API, so that our clients may exercise full control of their data and that of their users.