AI model building for data analysis in LEAs: A practical example
The use of artificial intelligence / machine learning models in law enforcement is becoming a necessity. A multitude of use cases range from video analytics of public surveillance systems with their facial recognition, object detection, and traffic safety control, to more specific pattern analysis of data used to uncover criminal enterprises. The common element is the analysis of the expanding amounts of available data to uncover key features, i.e. finding “needles in the haystack”.
The research project KISTRA has developed an application in collaboration with German LEAs to process a set of media data points consisting of images, videos and text using AI models for facial recognition, object detection and text analysis. The results are displayed in a unified user interface with a filterable map, highlighting the features of interest found. A unified user interface shows the results of the algorithms clearly and transparently. It focuses on functionalities that enable the user to further filter the dataset based on the objects, faces, and keywords in a configurable way. This enables the analyst to narrow down the dataset to select few data points of interest, and if needed to sort these data into different collections for export, further evaluation and integration with external tools.
We highlight the process and pitfalls of creating and deploying machine learning models and the need to curate non-biased, well-balanced training data, especially in light of the EU Artificial Intelligence Act proposal. The performance of the application is shown in a practical scenario, evaluating up to 100.000 media items with facial recognition, object detection and OCR and text analysis. Our use case shows that the time the analyst spends on finding relevant information can be reduced by more than 90% on average.
Co-author: Jens Elsner
