Should we trust Machine Learning based Bloodstain Pattern Analysis?

This thesis asks whether Machine Learning (ML) models used in Bloodstain Pattern Analysis (BPA) should be considered trustworthy in forensic contexts. While ML models often achieve high accuracy under standard evaluation, their suitability for high-stakes forensic use remains uncertain. Semi-structured interviews with BPA experts identified factors influencing trust, including consistency under input variation, the ability to explain or justify results, and transparency and accountability. Guided by these requirements, the thesis evaluates Random Forest and six additional classifiers, including Support Vector Machines, K-Nearest Neighbors, and Decision Trees, through sensitivity and robustness analyses. Sensitivity experiments examine how image resolution and feature selection affect performance, while robustness is assessed using adversarial perturbations. Despite strong accuracy under controlled conditions, the models exhibit instability under input variations and remain vulnerable to adversarial manipulation. These limitations suggest current ML models do not fully satisfy expert-defined trust requirements.