A Security-Preserving Ensemble Convolutional Neural Network Framework for Automated Forensic Image Evidence Classification

Syeda Naila Batool; Muhammad Yousif; Muhammad sarmad shakil; Ume Reem; Hina Bari

Syeda Naila Batool Govt Graduate College for Women, Dubai Mahal Road, Bahawalpur
Muhammad Yousif Lahore Garrison University
Muhammad sarmad shakil Department of Computer Science, Minhaj University Lahore, Pakistan
Ume Reem Department of Computer Science, Hajvery University, Lahore, Pakistan
Hina Bari School of Systems and Technology, Department of Informatics and Systems, University of Management and Technology Lahore

Keywords: Digital Forensics, Forensic Image Analysis, Ensemble Learning, Convolutional Neural Networks, Secure Deep Learning, Automated Evidence Classification

Abstract

Digital forensic investigations increasingly rely on image-based evidence collected from diverse and often unsecured sources, making accurate, automated, and secure analysis a critical requirement. This study proposes a security-preserving ensemble convolutional neural network (CNN) framework for the automated classification of forensic image evidence. The proposed system is designed to operate on images acquired from real-world digital forensic scenarios, including crime scenes, seized devices, and surveillance systems, where variations in lighting, noise, background complexity, and image quality are common. The framework utilizes an ensemble of transfer-learning-based CNN models to extract discriminative forensic features related to texture patterns, color distributions, structural anomalies, and object characteristics present in digital evidence. To address data sensitivity and integrity concerns inherent in forensic investigations, a security-preserving learning mechanism is incorporated to minimize data exposure while ensuring evidential reliability. Data augmentation techniques are employed to enhance robustness, mitigate overfitting, and handle class imbalance frequently observed in forensic datasets. The proposed system supports multi-class classification, enabling the identification of multiple categories of forensic image evidence with visually similar characteristics. Experimental evaluation demonstrates high classification accuracy and strong generalization performance across heterogeneous forensic datasets. The results indicate that the ensemble CNN framework effectively enhances the reliability, scalability, and security of automated forensic image analysis. This work contributes toward intelligent and secure digital forensic systems, supporting timely and informed decision-making during criminal investigations.