This talk will cover the technical challenges of leveraging the latest computer vision and machine learning techniques in a context of fraud detection. In particular, we will show how the latest advancements in deep learning allow us to significantly automate industrial processes that heavily rely on manual labeling. We will start with a presentation of our system leveraging biometrics and identity document data. We will present the key constraints of fraud detection, such as constantly evolving threats, massively unbalanced data, and unreliable labels. We will discuss how we learned to navigate them, while finding the best possible balance between false acceptance and false rejection. We will then move on to the trade-offs of supervised and unsupervised learning and how deep learning can be used most effectively in the fraud detection setting. Finally, we will focus on how we address bias in our system. Making identity verification as fair as possible is a core objective for Onfido. We will present concrete steps that practitioners can use in a real-world setting to reduce bias in their systems.