The Effect of Semantically Aligned Images for Deepfake Detection

Aritra Bandyopadhyay, Tyme Chatupanyachotikul, Jose Garcia, Arijit Ghosh, Karolina Hajkova, Carlos Patiño, Ivona Najdenkoska; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 1640-1649

Deepfake detection is crucial for ensuring security and trust in digital media, as it faces the ongoing challenge of methods that generalize across diverse generative models. Recent work has addressed this problem by identifying artifacts from the image generation process, such as Latent Reconstruction Error (LaRE), or with semantically aligned training datasets. Unfortunately, the two approaches applied separately do not achieve good generalization across all types of image generators. Our work combines LaRE features with a semantically aligned training dataset to investigate whether using both approaches enhances generalization in deepfake detection. We evaluate our approach using the GenImage dataset and find that, compared to a baseline with just LaRE features, our approach performs better for images generated with BigGAN and ADM but underperforms the baseline on average. We conduct ablation studies on each component of the process to provide insights into the complexities of this combination and identify critical factors for future improvement.