Ensuring Data Security and Annotators Anonymity Through a Secure and Anonymous Multiparty Annotation System

Healthcare is a domain characterized by the continuous influx of unannotated data generated by clinical workflows. The potential benefits of integrating artificial intelligence in support of clinical decisions are therefore substantial. An efficient machine learning approach should rely on a continual finetuning of the model with human expert annotations. In order to reduce the workload of the experts, Active Learning strategies are required. They involve the active participation of medical experts to provide the most informative annotations for model finetuning. The main goal of this paper is to propose an architecture which builds trust among the annotators in the process, thereby encouraging their engagement and motivation for ongoing participation in the annotation process. We provide a Secure and Anonymous Multiparty Annotation System (SAMAS) that could be used for securely annotating data in both Active Learning and Continuous annotation workflow. This architecture is based on conditional access mechanisms and communication scheme implying a Trusted Third Party to ensure both security of communications and annotators anonymity. We verified our implementation, and its compliance to a set of predefined properties, in a simplified example of distributed annotation using Spin and Linear Temporal Logic.