Virtual sample generation for few-shot source camera identification

The Source Camera Identification (SCI) has achieved remarkable success. However, existing approaches require sufficiently large training sets for high performance on accuracy and robustness. For maintaining high performance given small training sets, we propose a semi-supervised, Mega-Trent-Diffusion (MTD) method to generate virtual samples, such that the training sets can be expanded and unlabeled samples can be fully utilized as well. The stability of our method is improved using ensemble learning. Our theoretical analysis and experiments corroborate the effectiveness of our method beyond others when few-shot is given.