Reliable Camera Model Identification Through Uncertainty Estimation

Jiarong Pan, Anatol Maier, Benedikt Lorch, Christian Riess

Abstract: Camera model identification is a standard task in digital image forensics. Learning-based approaches achieve state-of-the-art performance, but they are sensitive to so-called out-of-distribution (OOD) data due to a mismatch between the training and testing distribution. This may result in a significant reduction in classifier performance that is, unfortunately, not easy to anticipate for a forensic analyst.In this work, we investigate possibilities for adding reliability measures to the task of camera model identification. We leverage learning architectures that include an uncertainty measure with every prediction that can be reported back to an analyst. To this end, we investigate deep ensembles and Bayesian neural networks (BNNs). We compare both methods against a standard CNN with softmax statistics as uncertainty metric. We demonstrate in several experiments that both probabilistic approaches provide simultaneously state-of-the-art classification performance and reliable uncertainty estimates on OOD data. The uncertainty of deep ensembles is more accurate on OOD camera models, while BNN uncertainties are more accurate on OOD post-processing.