A Deep Learning and Hybrid Optimization Framework for Digital Evidence Tamper Detection: Legal Reliability, Chain of Custody, and Cyber Forensics

Abstract

Digital evidence is an essential part of the toolbox in today's cybercrime investigations and court cases. The speed of progress of image editing software, the development of deepfake creation methods, however, and the introduction of anti-forensic tools have placed digital evidence under heightened pressure to be manipulated, rendering it questionable in terms of authenticity, integrity, and admissibility in court. Despite the recent deep learning methods that showed promising results in digital image forgery detection, existing methods mainly emphasize classification accuracy without offering satisfactory interpretability of the model and they are not sufficiently supported by forensic reliability. To tackle these difficulties, the present paper introduces a novel deep learning and hybrid optimization approach for detecting the tampering in digital evidence that combines Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Particle Swarm Optimization–Grey Wolf Optimizer (PSO–GWO), Explainable Artificial Intelligence (XAI), and cryptographic evidence integrity verification. First, digital evidence is preprocessed to retain the forensics artifacts, and then, the complementary local and global features are extracted using the hybrid CNN–ViT architecture. The extracted features are then enhanced by using PSO–GWO for the optimization of the features to maximize discriminative capability before classification. Additionally, Grad-CAM and SHapley Additive exPlanations (SHAP) are integrated to visually and feature-wise explain the model's decisions, while the SHA-256-based hash verification helps maintain the integrity of the evidence within the forensic workflow. Experimental results show that the proposed framework is able to detect the manipulated digital evidence with overall accuracy, precision, recall, F1 score, and ROC-AUC of 98.4%, 98.1%, 98.0%, 98.0%, and 0.992 respectively, respectively. The proposed architecture combines all the elements of a reliable AI-assisted digital forensic investigation in a single framework, offering a transparent and reliable solution for such an investigation.