Anti-noise image source identification

Image source identification plays an important role in image forensics. Traditional image source identification techniques rely on the assumption that the training datasets are well labeled. However, in the real world, this assumption may not be the truth, and some noisy samples may exist in the training datasets. In this paper, firstly, we theoretically investigate the influence of noisy samples to the identification performance. Then, a new image source identification approach, namely, Anti-noise Image Source Identification (AISI), is proposed to deal with those noisy samples. AISI has three steps, ie, noise level evaluation, noise level based sampling, and multi-classification. Noise level evaluation aims to assign each sample with a noise level that indicates the probability of being noisy. The basic idea of noise level based sampling is to sample images according to their noise levels. We provide a theoretically justification to demonstrate the effectiveness of AISI. Experiments conducted on a real-world image collection confirm that the proposed AISI can alleviate the influence of noisy samples and can improve the identification accuracy while noise exists.