Maximize margins for robust splicing detection

TL;DR

This paper investigates the variability of deep learning-based splicing detectors under post-processing and proposes a training strategy to enhance their robustness by maximizing latent margins.

Contribution

It reveals the link between latent space margins and detector robustness, and introduces a method to improve generalization by training multiple variants and selecting the best.

Findings

Latent margin distribution correlates with robustness to post-processing.

Training multiple models under different conditions improves detection reliability.

Maximizing latent margins enhances generalization to unseen post-processed images.

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.