arXiv:2508.00897v1 Announce Type: cross Abstract: Despite recent progress in splicing detection, deep learning-based forensic tools remain difficult to deploy in practice due to their high sensitivity to training conditions. Even mild post-processing applied to evaluation images can significantly degrade detector performance, raising concerns about their reliability in operational contexts. In this work, we show that the same deep architecture can react very differently to unseen post-processing depending on the learned weights, despite achieving similar accuracy on in-distribution test data. This variability stems from differences in the latent spaces induced by training, which affect how samples are separated internally. Our experiments reveal a strong correlation between the distribution of latent margins and a detector's ability to generalize to post-processed images. Based on this observation, we propose a practical strategy for building more robust detectors: train several variants of the same model under different conditions, and select the one that maximizes latent margins.