Appen Launches New LLM Features for Enterprises.

Onfido's Search for a Flexible On-Premise Labeling Tool

How Onfido's labeling tool enhanced machine learning
March 15, 2023
Share

The Company

Onfido is a global provider of identity verification. The company makes it easy for end-users to prove their identities online. So they can access services, products and businesses online. Their Real Identity Platform allows for biometric and identity document verification, with fraud detection built in. It works on desktops and mobile devices like cell phones, for situations when online convenience is preferred over face-to-face authentication.

Since the beginning of the pandemic, identify fraud has been up 41% with increases in sophisticated techniques fraudsters used such as 3D silicone masks, display attacks, and high-quality printouts. Onfido’s goal is to help businesses stop fraud and reduce customer acquisition costs in the process.

The Challenge

Inherent to Onfido’s work with data are significant security challenges. Dealing with biometrics, most data Onfido processes contain personally identifiable information (PII). Every time they share data with a third party, they have to get special permission from each company they work with to do so.

Onfido needed a labeling tool that could be installed on-premises, that could be flexibly configured and had powerful annotation capabilities across a wide variety of data types including images, videos, and documents.

The Solution

In their search to find a labelling partner, Onfido came across two challenges. First, they needed to find a company that offered both an on-premise deployment option and provided the ability to customize the tooling used to label the kind of data they needed. Onfido found the request of labeling tool customization to be a demanding ask, but they knew Appen would be able to do it and meet both their requirements for a labelling partner.

Our dedicated team met weekly to discuss the labeling process and work together to build a model that would provide the precise training data Onfido needed to properly train their machine learning model. After several meetings we were able to create the exact product they needed for data labeling. Appen’s collaboration with Onfido provided a customized experience that they could scale and execute.

Onfido took advantage of Appen’s breadth and depth as a labeling partner by leveraging Appen’s data collection services to source identity data according to specific demographic requirements.

“We appreciate all the support, from the customized project plan to the weekly status meetings, Appen is always there for us.”  – Francois Jehl, Director of Data and Analytics

The Result

Our data labeling tool was vital in training the machine learning models that power their product, Motion.  It leverages anti-fraud technology and a simple UX design that just needs a user to do simple head turns for verification approval. To train Motion’s anti-fraud models, Onfido launched a dedicated lab to generate hundreds of thousands of fraud videos using masks, mannequins, and videos of videos, and labeled them using Appen’s tool which allowed them to improve the performance of their anti-fraud models by 10x.

Onfido built a continuous human-in-the-loop retraining pipeline for a production model. It measures image quality using Appen tooling. It takes a subset of production data, automatically generates annotation jobs using the Appen API, gets some of the data manually labeled, and sends it back to retrain the models.

"In a nutshell, Appen was paramount in the development of Onfido and has become a critical component of our operations. We are thrilled to have found a reliable partner in Appen." – Francois Jehl, Director of Data and Analytics at Onfido.

See our partnership in action with a demo of Onfido’s biometric security in this video from BBC.