OneDetect: A Federated Learning Architecture for Global Soft Biometrics Prediction
Bilal Hassan, Ebroul Izquierdo
Abstract
Following the amount of threats and increased number of terrorism reported over the years, the development of security technologies is becoming a crucial subject, in recent era. Moreover, the paradigm of security technologies is shifting from concept of multiple representations to one and unique identity for each individual - like there are several different identity verification systems exist in the world, developed by different countries. On the other hand, security is a broad subject with its applications at open places like markets and roads and verification during online sessions, etc., however, border control is one of the most important point. Right now, biometrics like finger prints and facial scans are most widely used technology, termed as traditional or non-intrusive biometrics. To provide a single representation based unique identity verification system, we are going to propose a federated learning architecture called OneDetect, which uses intrusive features from human body to predict three most common global soft biometrics, e.g., gender, age and ethnicity. The purpose of using federated architecture is to ensure privacy of data from each client, i.e., country or region, while developing a more generalized verification system where each verified instance improves the verification system overall. Additionally, the use of global soft biometrics in such kind of architecture will provide seamless recognition whether, it is an airport, seaport or any public place. To achieve this goal, we recorded our own dataset called MMV Pedestrian dataset, in an airport like walking corridor and EfficientNetB3 architecture is used for training and prediction by each client.