Jumio advises on dealing with biometric bias in online ID verifications
Jumio has been explaining the risk of bias embedded in biometric machine learning models and Artificial Intelligence-based systems, and is suggesting practical methods through which such biases can be reduced in order to minimize the impact on people’s daily lives.
In a piece titled “5 practical ways to reduce AI bias in online identity verification,” Jumio also examined the various forms of AI bias and the demographic traits through which they manifest. The piece underlines facial recognition processes as being one of the most liable to biases of a racial nature.
According to Jumio, customers should be able to ask questions to would-be service providers in order to have a better understanding of how they are addressing demographic bias.
The five things companies should find out about from biometrics providers are the size and representativeness of the provider’s biometric training database, where the data creating the training data sets come from, how the data sets are labelled, the type of quality controls governing the tagging process, and the diversity of the team developing the algorithms, Jumio says.
The blog post states that if all of these are taken into consideration, bias which is often inadvertent in AI algorithms, will be significantly curbed. It also cites the 2020 Gartner Market Guide for Identity Proofing and Affirmation which suggests that “by 2022, more than 95 percent of RFPs for document-centric identity proofing will contain clear requirements regarding minimizing demographic bias, an increase from fewer than 15 percent today.”
Jumio posits that many leading identity verification solutions leverage AI and machine learning biometrics to assess the digital identity of remote users — and, unfortunately, these algorithms are also susceptible to demographic bias which include race, age, gender and other characteristics.
Jumio notes that although demographic bias is for the most part an unconscious act because many facial recognition solution providers do not necessarily know when they are making the algorithm that it is likely going to produce incorrect outcomes at uneven rates, the impact of bias on a company can be very far-reaching as it could taint their image or even provoke litigations from users.
Meanwhile, Jumio says it is making efforts in its daily work in order to address concerns regarding its AI algorithms and processes. In this regard, it has elaborated a new reference guide to achieve this goal and this includes what it says is its large and representative data sets, production of realistic data, and quality control and governance, among others.
accuracy | AI | algorithms | biometric identification | biometric-bias | biometrics | digital identity | facial recognition | identity verification | training