A Comparative Analysis of PRNU-Based Methods and Convolutional Neural Networks for Source Camera Identification

Source camera identification is a critical digital forensics process, whose usage extends to legal investigations, cybersecurity, and multimedia authentication. The classical algorithms based on Photo Response Non-Uniformity have been widely used due to their ability of utilizing different patterns of sensor noise for source identification. The advent of deep learning, in particular Convolutional Neural Networks, however, introduced new opportunities for improved accuracy and stability in difficult cases. This paper gives an in-depth comparison of classical PRNU-based algorithms and algorithms based on Convolutional Neural Networks for source camera identification. We evaluate these algorithms on a diverse set of images captured using different camera models, comparing their performances in terms of accuracy, computation, and stability in the face of standard image manipulation such as resampling, adding noise, and compression. Our experiments demonstrate that though computationally powerful and highly effective in ideal cases, in manipulated and noisy images, CNN is a better alternative, producing improved results in terms of accuracy and generalization. This paper is highlighting the trade-off between classical and deep learning solutions, which can be of benefit for digital forensics practitioners and scientists. The outcome of this study shows that a combination of the best of PRNU and CNN can be a valuable direction for further study in source camera identification.