Nadia Siddiqui

Camera model identification remains a core challenge in digital image forensics for source attribution, integrity verification, and copyright protection. While rich model residuals effectively capture sensor and in-camera processing artifacts, most existing approaches rely on the high-dimensional residual set, leading to redundancy, high computational cost, and limited exploration of large-scale multi-class settings. This work introduces a systematic feature-engineering framework that selectively retains only the most discriminative residual components, reducing dimensionality from 34,671 to 7,176 features. This approach significantly enhances computational efficiency and class separability while preserving forensic performance. The method is explicitly optimized for scalable multi-class camera model identification, providing a lightweight yet robust alternative to both traditional and deep-learning-based techniques. Extensive experiments were conducted on the Dresden and VISION datasets across inter-brand, intra-brand, and large-scale multi-class scenarios. On 18 and 23 camera models from Dresden and 29 camera models from VISION, the proposed framework achieved classification accuracies of 98.22%, 97.12% and 96.45%, respectively, demonstrating strong scalability and generalization. Robustness was further validated under realistic post-processing transformations, including JPEG compression (QF = 50 and QF = 80), resizing, and center cropping. The model maintained high diagonal dominance in confusion matrices under these degradations, confirming resilience to common real-world image manipulations. Overall, the selective residual modeling combined with non-linear kernel-based classification delivers a novel, computationally efficient, and high-performing solution for large-scale camera model identification. It outperforms or matches recent deep learning approaches while requiring significantly fewer computational resources, making it particularly suitable for practical forensic applications.