Less is More: Deep Learning Framework for Digital Forensics in Resource-Constrained Environments

Machine learning (ML)-enabled tools and techniques are an unavoidable necessity for modern digital forensics. However, the performance of such tools often requires extensive training using powerful computing environments that may expend thousands of dollars and several training hours for a single model. Further, this process necessitates using high-volume datasets that are generally unavailable due to court-mandated restrictions. To overcome these problems, this paper proposes using a two-layered strategy that fuses several features of transfer learning and knowledge distillation-based techniques to build lightweight neural networks that can be trained using low-volume datasets to achieve high accuracy. Experimental results highlight we achieved an accuracy of 98.5% and 97.5% during training and testing of the heavier ML model, and an accuracy of 97.08% and 97.2% on the lightweight ML model (during training and testing respectively), therefore demonstrating the robustness and validity of the proposed technique.