<p>The objective of source camera identification (SCI) is to ascertain the originating device of target images, ensuring the reliability of digital image sources. However, current state-of-the-art techniques often rely on an abundant supply of training samples, which proves challenging to acquire in practical settings. To solve it, we propose a novel approach called residual information distillation (RID) to address the issue of few-shot sample databases. By leveraging the concept of knowledge transfer across datasets, our method optimizes model performance in scenarios with a limited number of training samples. The proposed approach involves training a teacher model using publicly available datasets that offer a wealth of samples. Subsequently, the knowledge encapsulated in the teacher model’s weights is distilled into a student model, which is trained on few-shot sample datasets, to facilitate more effective training. Extensive experiments on Dresden, VISION, and SOCRatES datasets show that RID outperforms state-of-the-art methods. In the 10-shot setting, our method achieves 89.07%, 85.64%, and 79.08% accuracy, with notable improvements over existing approaches. It also maintains strong performance in 5-shot, cross-network, and cross-task distillation scenarios. The proposed RID effectively alleviates overfitting and improves generalization, making it practical for real-world few-shot forensic applications.</p>