MUN: Image Forgery Localization Based on M³ Encoder and UN Decoder
Authors: Yaqi Liu, Shuhuan Chen, Haichao Shi, Xiao-Yu Zhang, Song Xiao, Qiang Cai
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on publicly available datasets show that MUN outperforms the state-of-the-art works. Experiments Implementation and Evaluation Details Experimental datasets. Experimental metric. Ablation Study Four groups of experiments for ablation study are conducted on the testing set of the Synthesized Dataset. Robustness Evaluation In practice, one may disguise forged images with additional postprocessing. Comparison with State-of-the-art Methods We compare MUN with recent SOTA works, including RGB-N (Zhou et al. 2018), Man Tra-Net (Wu, Abd Almageed, and Natarajan 2019), SPAN (Hu et al. 2020), MVSS-Net (Chen et al. 2021), PSCC-Net (Liu et al. 2022a), Object Former (Wang et al. 2022), TANet (Shi, Chen, and Zhang 2023), TBFormer (Liu et al. 2023), Hi Fi (Guo et al. 2023), Tru For (Guillaro et al. 2023), CSR-Net (Zhang et al. 2024), NRL-Net (Zhu et al. 2024) and MGQFormer (Zeng et al. 2024). Table 6 shows the results of above-mentioned models on NIST16, CAISA v1.0, IMD2020, Coco Glide and Wild datasets, and the scores are borrowed from their original papers. |
| Researcher Affiliation | Academia | Yaqi Liu1*, Shuhuan Chen2,3*, Haichao Shi2, Xiao-Yu Zhang2 , Song Xiao1, Qiang Cai4 1Beijing Electronic Science and Technology Institute 2Institute of Information Engineering, Chinese Academy of Sciences 3School of Cyber Security, University of Chinese Academy of Sciences 4Beijing Key Laboratory of Big Data Technology for Food Safety, Beijing Technology and Business University EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology with equations and textual descriptions (e.g., in sections 'M3 Encoder', 'UN Decoder', 'IoUDCE Loss', 'Deviation Noise Augmentation') but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/MrHuan3/MUN |
| Open Datasets | Yes | We train our model on the Synthesized Dataset in (Liu et al. 2023) and adopt five datasets, i.e., NIST16 (Guan et al. 2019), CASIA v1.0 (Dong, Wang, and Tan 2013), IMD2020 (Novoz amsk y, Mahdian, and Saic 2020), Coco Glide (Guillaro et al. 2023), Wild (Huh et al. 2018), for evaluation. |
| Dataset Splits | Yes | There are 156006 synthesized images in the Synthesized Dataset (140432 for training, 7787 for validation, and 7787 for testing). |
| Hardware Specification | Yes | Our training and inference experiments are conducted on a single NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | No | MUN is built based on MMSegmentation. The paper mentions software tools like MMSegmentation but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We use Noiseprint++ to generate the noise map of the original-sized image and then resize both the image and the noise map to 512 512 patches as inputs. The optimizer is SGD, and the learning strategy can be calculated as follows: ( lrs + (lr0 lrs) iterc iterw , iterc < iterw lr0 (1 1 itert iterc+1), iterc iterw (13) where lrc denotes the current learning rate. lrs = 10 6 is the start learning rate and lr0 = 0.01 is the initial learning rate, iterc stands for the current number of iterations, iterw = 1500 denotes the warm-up number of iterations and itert means the total number of iterations. In Eq. (9), α0, α1 are both set to 1.0 and λ is set to 1.5. The batch size is 7 and the epoch is 4. |