Computer Vision News - October 2016

fake documents, they report it to police as well. This doesn’t happen in the USA however, since they don’t share this information, which according to Kassai is a big shame. Such exchanges are much better accepted in Europe. Of course, there are different levels of fake documents: poor quality fakes which are regularly caught; moderate quality fakes which are sometimes caught; and sophisticated fakes which the client reports to Onfido, giving them precious feedback for the future. Essentially, the more checks they do, the more robust the app becomes, which is why technology and machine learning works better than with humans. It is so much more robust. The specific algorithms that help solve these challenges have come about after four years of research, during which they tried different techniques and found the ones that work the best. The technology uses an ensemble of techniques to get the needed results, ranging from what would be considered more classical machine learning and computer vision techniques (SVM’s, feature extraction and the like) to the use of deep learning and a combination of these methods. The nature of their work, Kassai explains, often means that they are dealing with a considerable amount of uncertainty and they are often balancing this risk and reward with the techniques that they are using. Their goal in doing that is to create an end-to-end trainable system. However, to achieve that goal they had to break the problem into a number of smaller problems which they are iterating over and then pulling these solutions together to create the end-to-end approach. This consolidation process may sometimes be quite painful: the team may throw away weeks or months of work and fine tuning to replace it with a more generalised model or approach, however Onfido feels that it is all part of the process, when the system is constantly evolving and growing. Onfido makes use of a huge variation of techniques, including convolutional networks, LSTMs, SVMs and many more. Privacy is a very big concern with all identity verification systems. When asked about what happens if an individual is concerned with exposing his or her own personal documents, Kassai replies: “There are two things that happen. One is that we only check all persons with their consent. When you sign up with an online bank, a remittance platform, or with any client of ours, we first take your consent before we are able to verify you. “The second thing is that most people understand that in order to open a bank account, you have to show up to the bank, queue up and show your details. It takes a long time. Whereas, if you want to do it online, then there needs to be a robust way to do it. Otherwise there would be so much 22 Computer Vision News Application Application Machine learning engineers at Onfido