Towards Federated Learning Based Contraband Detection Within Airport Baggage X-Rays

To maintain border and transportation security against a variety of threat profiles, it is essential to find contraband in airport baggage X-rays. Security inspectors have a tougher time avoiding misdetection due to a lack of support staff and a stressful workplace. Although machine learning models can detect contraband automatically, most of the models used in this process train the data in a centralized learning (CL) way, posing possible security and privacy concerns. To address this, we propose a Federated learning (FL)-based architecture to detect contraband in x-ray baggage security images while maintaining user privacy. Our model is trained and evaluated using the most recent state-of-the-art YOLOv7, SSD, and Faster R-CNN algorithms, paving the door for large-scale automatic detection of contraband in airports across the globe through collaboration. We achieve a global accuracy of 90.1%, 86.4%, and 66.7% with Faster R-CNN, SSD, and YOLOv7 algorithms respectively using the PIDray dataset. Our experiment reveals the challenges and potential of utilizing FL to detect contraband in airport luggage X-rays, even though the performance is comparable with the benchmark performance of non-FL algorithms.