Federated tool for anonymization and annotation in image data

The increasing complexity of security challenges requires Law Enforcement Agencies (LEAs) to have improved analysis capabilities, e.g., with the use of Artificial Intelligence (AI). However, it is challenging to make large enough high-quality training and testing datasets available to the community that is developing AI tools to support LEAs in their daily work. Due to legal and ethical issues, it is often undesirable to share raw data with personal information. These issues can lead to a chicken-egg problem, where annotation/anonymization and development of an AI tool depend on each other. This paper presents a federated tool for semi-automatic anonymization and annotation that facilitates the sharing of AI models and anonymized data without sharing raw data with personal information. The tool uses federated learning to jointly train object detection models to reach higher performance by combining the annotation efforts of multiple organizations. These models are used to assist a person to anonymize or annotate image data more efficiently with human oversight. The results show that our privacy-enhancing federated approach – where only models are shared – is almost as good as a centralized approach with access to all data.