Search

Various fields have benefited from machine learning (ML) applications (e.g., health, finance, military), mainly for automating tasks or improving decision making capability. The area of digital forensics shows potential promise as categorizing content of seized devices for potential evidentiary value is a complicated task due to the increasing amount of data that needs to be processed. A rapid adaptation of ML-based products for such a real-life application requires the creation of engineering knowledge that would make ML models more accessible for developers while incorporating human-centric validation into user acceptance procedures. This study proposes a development process to incorporate pre-trained models to the development life-cycle of an ML-based digital forensic tool that classifies images. We applied the process and created a prototype tool that identifies handguns by rigorously following the design science methodology. We evaluated four ImageNet-trained models: InceptionV3, Xception, ResNet and VGG16. Using realistic datasets and various decision criteria, we selected the outperforming model and utilized it in the tool. The usability and learnability of the prototype were measured using the System Usability Scale (SUS) questionnaire as an early-phase usability test. Our tool was able to achieve an acceptable rating in the SUS results from the average of five respondents. This study successfully demonstrated that freely available pre-trained machine learning models without fine-tuning are feasible for use in practice.