Reliable JPEG Forensics via Model Uncertainty

Many methods in image forensics are sensitive to varying amounts of JPEG compression. To mitigate this issue, it is either possible to a) build detectors that better generalize to unknown JPEG settings, or to b) train multiple detectors, where each is specialized to a narrow range of JPEG qualities. While the first approach is currently an open challenge, the second approach may silently fail, even for only slight mismatches in training and testing distributions. To alleviate this challenge, we propose a forensic detector that is able to express uncertainty in its predictions. This allows detecting test samples for which the training distribution is not representative. More specifically, we propose Bayesian logistic regression as an instance of an infinite ensemble of classifiers. The ensemble agrees in its predictions from test samples similar to the training data but its predictions diverge for unknown test samples. The applicability of the proposed method is evaluated on the task of detecting JPEG double compression. The detector achieves high performance on two goals simultaneously: It accurately detects double-JPEG compression, and it accurately indicates when the test data is not covered by the training data. We assert that the proposed method can assist a forensic analyst in assessing detector reliability and in anticipating failure cases for specific inputs.