Learnable Frequency Decomposition for Image Forgery Detection and Localization

Dong Li, Jiaying Zhu, Yidi Liu, Xin Lu, Xueyang Fu, Jiawei Liu, Aiping Liu, Zheng-Jun Zha

Introduction

Image forgery detection and localization (IFDL) has become an increasingly difficult challenge as sophisticated editing tools make forged regions nearly indistinguishable from authentic ones in the standard RGB spatial domain. While many forensic methods have turned to the frequency domain, they often treat it as a secondary feature or focus exclusively on high-frequency noise. Researchers from the University of Science and Technology of China have introduced F2D-Net, a framework that adaptively decomposes images into phase and amplitude components to capture subtle manipulation traces across the entire frequency spectrum.

The F2D-Net Architecture

The Forensic Frequency Decomposition Network (F2D-Net) is designed to overcome the limitations of fixed-filter frequency analysis. It operates in two primary stages:

  1. Spectral Decomposition Subnetwork (SDSN): This module utilizes deep Fourier transforms to split an image into its amplitude and phase components. It incorporates Phase-Emphasized Interaction (PEI) blocks to prioritize learning from the phase spectrum, which is critical for identifying forged objects. To maintain spatial context, Space-Frequency Dual Cross Attention (SFDCA) is used to bridge the gap between frequency analysis and spatial localization.
  2. Frequency Separation Subnetwork (FSSN): Unlike traditional methods that use static frequency cutoffs, the FSSN adaptively predicts a mask to separate the SDSN's output into high and low-frequency bands. This "divide-and-conquer" strategy allows the model to refine each band independently before fusing them into a final localization map.
The architecture of F2D-Net, featuring the Spectral Decomposition and Frequency Separation subnetworks.

The Importance of Phase and Multi-Band Analysis

A key insight of this research is that forgery traces are not confined to high-frequency noise. Statistical analysis reveals that manipulation residuals are significant in both low-low (LL) and high-high (HH) frequency bands. Furthermore, the authors demonstrate that the phase component is a "smoking gun" for forensics; while amplitude typically captures style and texture, phase stores the structural and semantic information of an image. Subtracting the phase spectra of real and forged images reveals manipulation traces far more clearly than spatial RGB residuals.

Motivation for F2D-Net: (a) Comparison of RGB residuals vs. phase and amplitude residuals; (b) Distribution of forgery traces across frequency bands.

Experimental Results

F2D-Net was pre-trained on a large-scale synthetic dataset and evaluated against several benchmarks, including NIST16, CASIA, and IMD20. The model achieved state-of-the-art (SOTA) performance:

  • Localization Precision: On the NIST16 dataset, it reached a pixel-level AUC of 99.8% and an F1 score of 87.2%, outperforming previous leaders like PSCCNet.
  • Detection Accuracy: On the CASIA-D dataset, it achieved a detection AUC of 99.73%.
  • Robustness: The model remains highly effective even under heavy JPEG compression (quality factor = 50) or significant Gaussian noise, where competing models often fail.

Conclusion

By treating frequency decomposition as a learnable process rather than a fixed pre-processing step, F2D-Net effectively uncovers forgery traces that are invisible in the spatial domain. While the model shows slight performance variations on specific splicing styles (e.g., the Columbia dataset), its ability to leverage phase information and adaptive frequency separation sets a new standard for robust image forensics. Future work may focus on optimizing the computational complexity of Fourier transforms and extending the model to detect purely AI-generated content.

Summary updated

Abstract

Concern for image authenticity spurs research in image forgery detection and localization (IFDL). Most deep learning-based methods focus primarily on spatial domain modeling and have not fully explored frequency domain strategies. In this paper, we observe and analyze the frequency characteristic changes caused by image tampering. Observations indicate that manipulation traces are especially prominent in phase components and span both low and high-frequency bands. Based on these findings, we propose a forensic frequency decomposition network (F2D-Net), which incorporates deep Fourier transforms and leverages both phase information and high and low-frequency components to enhance IFDL. Specifically, F2D-Net consists of the Spectral Decomposition Subnetwork (SDSN) and the Frequency Separation Subnetwork (FSSN). The former decomposes the image into amplitude and phase, focusing on learning the semantic content in the phase spectrum to identify forged objects, thus improving forgery detection accuracy. The latter further adaptively decomposes the output of the SDSN to obtain corresponding high and low frequencies, and applies a divide-and-conquer strategy to refine each frequency band, mitigating the optimization difficulties caused by coupled forgery traces across different frequencies, thereby better capturing the pixels belonging to the forged object to improve localization accuracy. Experiments on multiple datasets demonstrate that our method outperforms state-of-the-art image forgery detection and localization techniques both qualitatively and quantitatively.
BibTeX
@inproceedings{li2025learnable,
  author = {Dong Li and Jiaying Zhu and Yidi Liu and Xin Lu and Xueyang Fu and Jiawei Liu and Aiping Liu and Zheng-Jun Zha},
  title = {{Learnable Frequency Decomposition for Image Forgery Detection and Localization}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/152},
}
Figure