Our paper titled “Ignoring the Decoy: Exposing and Tackling Forensic Distractions in Image Forgery Localization using Masked Convolutions” was accepted and presented at the Synthetic Realities and Data in Biometrics and Security Workshop (SynRDinBAS) at the Winter Conference on Applications in Computer Vision (WACV) 2026 in Tucson, Arizona, USA.
In this paper, we investigated the impact of forensic distractions on Image Forgery Localization (IFL) methods. IFL methods aim to localize manipulated regions in images. However, we found that innocent visual elements, such as logos, captions or blurred faces, can act as decoys that mislead IFL methods. They draw attention away from the actual manipulated regions, leading to incorrect localization results.
A series of feature maps extracted from TruFor. Throughout the convolutional layers, the logo attracts more and more attention, while the actual manipulation (Zelenskyy and his wife) disappears to the background.
We propose a novel adaption of the convolution operation, called masked convolutions. By masking out the distracting regions (manually or automatically selected) during the convolution operations, our method allows existing IFL methods to ignore the decoys and refocus on the actual manipulated regions. Masked convolutions can be easily integrated into existing CNN-based IFL methods, without requiring retraining or additional data.
We evaluated our method on two sensitive state-of-the-art IFL models, CAT-Net and TruFor experiencing performance drops of 15% and 56% respectively. With our masked convolution approach, the drops were reduced to just 4% and 3%, demonstrating the potential of our method to improve the robustness of IFL methods against forensic distractions.
The paper was presented at the SynRDinBAS workshop at WACV 2026, on 6 March 2026, in Tucson, Arizona, USA, by Xander Staelens.
