This experiment developed and tested a machine learning algorithm trained on synthetic data to support investigators to identify and analyse UK-manufactured cluster munitions in footage of airstrikes. The researchers hypothesised that the tool could identify cluster munitions in videos more efficiently than manual filtering. The experiment also developed a secure way to store and manage evidence collections which is crucial as images and videos containing possible evidence are often quickly removed either by the platform or the poster.
The experiment is part of an ongoing investigation and we are still awaiting final conclusions. Results to date suggest that machine learning tools can be successfully developed and applied to support humans in data analysis, even where there is limited ground truth data by using synthetic datasets for training. However, expert verification of videos flagged by the algorithm as containing cluster munitions is still necessary. In addition, the object detector is still undergoing benchmarking for accuracy, so while using a machine learning tool to identify potential evidence saves time, the precision of the algorithm is still unclear.
Technologies like smartphones, social media, and participatory mapping tools enable witnesses to human rights violations to capture and share footage of war crimes in real time. These images create a wealth of potential evidence for legal investigations. However, manually reviewing the data is time consuming, expensive, and can cause trauma to investigators due to the disturbing nature of the content of the footage. In addition, the willingness of civilian witnesses or organisations to share this data depends on the ability to handle sensitive information securely. For courts to afford the information weight, it is also crucial to demonstrate that the evidence hasn’t been tampered with.
Follow the project lead on Twitter: @ProfYvo