SFNet: Spatio-Frequency Forgery Detection

• SFNet is a remote sensing image forgery detection framework that fuses spatial and frequency-domain features to enhance generalization across diverse terrains and generative models.
• It employs two independent feature extractors with Domain Feature Mapping and CBAM attention to align, fuse, and refine discriminative cues before binary classification.
• Empirical evaluations report a 4%–15.18% accuracy improvement over state-of-the-art methods across multiple datasets, demonstrating robust cross-domain performance.

Spatio-Frequency Network (SFNet) is a remote sensing image forgery detection framework that jointly models spatial-domain and frequency-domain evidence in order to detect fake remote sensing imagery across diverse terrains, land-cover types, and generative model families. It was introduced to address the limited generalization of single-cue detectors, which either overfit to spatial semantics or rely too narrowly on frequency artifacts tied to particular generators. SFNet uses two independent feature extractors, aligns their outputs in a common subspace through Domain Feature Mapping, fuses them by addition, and refines the result with a residual CNN and CBAM attention before binary classification. On three datasets, it is reported to achieve an accuracy improvement of 4%–15.18% over state-of-the-art remote sensing forgery detection methods while exhibiting robust cross-domain generalization (Qi et al., 25 Jun 2025).

1. Problem setting and design objective

SFNet was proposed in the context of remote sensing image forgery detection, a task made increasingly difficult by advances in GANs and diffusion models. The operating regime is unusually challenging because remote sensing imagery spans wide geographic coverage, diverse land-cover categories such as deserts, oceans, and residential areas, variable resolutions, and complex imaging or preprocessing effects including atmospheric interference and ortho-rectification. These factors can introduce artifacts in real images that resemble fake signatures, thereby weakening detectors that rely on narrow forensic priors (Qi et al., 25 Jun 2025).

The motivating critique is directed at single-cue detection. Spatial-only detectors often overfit to content semantics and prominent objects such as roads, buildings, or ships. Because high-level semantic features are shared between real and fake imagery, such detectors can miss subtle forensic traces. Frequency-only detectors, by contrast, exploit upsampling artifacts characteristic of GANs, but these artifacts vary across model families and become increasingly subtle in diffusion outputs. A detector that concentrates only on frequency priors may therefore generalize poorly across terrains and generators. SFNet is explicitly framed as a response to this failure mode: it seeks robust generalization by leveraging complementary spatial-domain and frequency-domain features while suppressing redundancy and overfitting.

A common misconception is that one domain should dominate the other. The reported formulation argues instead that the discriminative artifacts themselves change with geographic terrain, land cover, and generator type. This suggests that the central design problem is not simply stronger feature extraction, but domain alignment and fusion under distribution shift.

2. Architectural composition and feature extraction

SFNet operates on 3-channel remote sensing images resized to $256 \times 256$ and normalized. The input may be RGB or NIR-R-G depending on the dataset. Its front end consists of two independent feature extractors. The spatial-domain extractor, also termed the Image Extractor, is deliberately shallow so that it prioritizes local, low-level spatial cues rather than high-level semantics. It is defined as

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$

The implementation uses a small $C_s$ to emphasize low-level features rather than semantic abstraction (Qi et al., 25 Jun 2025).

The frequency-domain extractor is more elaborate and is built from HFRI, HFRF-S/HFRF-C, and FCL. HFRI, or High-Frequency Representation of Image, decouples spatial and frequency content, applies a high-pass filter, and reconstructs to the spatial domain:

$x' = \mathrm{iFFT}(B(\mathrm{FFT}(x))).$

The high-pass mask removes a centered low-frequency band:

$$B(f_{i,j}) = \begin{cases} 0, & |i| < W_f/4 \text{ and } |j| < H_f/4 \ f_{i,j}, & \text{otherwise.} \end{cases}$$

This suppresses low frequencies and emphasizes high-frequency edges and textures. HFRF-S and HFRF-C then enhance high-frequency content across the spatial or channel dimensions of intermediate features:

$$f'(\mathrm{dim}) = \mathrm{Conv}(\mathrm{iFFT}_{\mathrm{dim}}(B(\mathrm{FFT}_{\mathrm{dim}}(f)))).$$

The final Frequency Conv Layer performs learnable spectral filtering over amplitude and phase:

$$\tilde{f} = \mathrm{FFT}(\mathrm{Conv}_1(f)), \qquad \hat{f} = \mathrm{Conv}_{\mathrm{phase}}(\tilde{f}) + \mathrm{Conv}_{\mathrm{amplitude}}(\tilde{f}), \qquad F = \mathrm{Conv}_2(\mathrm{iFFT}(\hat{f})).$$

The resulting frequency representation is $F \in \mathbb{R}^{C_f \times H \times W}$.

The underlying Fourier transform is applied channel-wise, and the paper also states the standard FFT and inverse FFT definitions for a monochrome image $I(x,y)$:

$$F(u,v) = \sum_{x=0}^{H-1}\sum_{y=0}^{W-1} I(x,y)e^{-j2\pi(ux/H+vy/W)},$$

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$0

A DCT formulation is also given, but it is used only for dataset visualization and not in the SFNet forward path.

3. Domain alignment, fusion, and decision mechanism

After the two extractors produce $$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$1 and $$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$2, SFNet aligns them with a Domain Feature Mapping module. The purpose of DFM is to map the two feature distributions into a common subspace while, in the paper’s phrasing, “seeking commonality and preserving differences.” The aligned features are

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$3

where each projection function is composed of three stacked projection blocks. A projection block is defined as

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$4

followed by a residual addition,

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$5

The spatial and frequency branches do not share parameters. The residual shortcuts are intended to preserve key low-level forensic cues during nonlinear transformation (Qi et al., 25 Jun 2025).

Fusion is performed by element-wise addition rather than naive concatenation:

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$6

An optional batch normalization step may follow. The choice of addition is motivated by the claim that naive concatenation can destabilize fusion weights when the two domains contain semantic inconsistencies. DFM is therefore placed before fusion to alleviate distribution shift.

The fused representation is then refined by a residual CNN backbone and CBAM attention. Channel attention is applied first, then spatial attention:

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$7

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$8

$$S = \mathrm{MaxPool}(\mathrm{ReLU}(\mathrm{BN}(\mathrm{Conv}(x)))) \in \mathbb{R}^{C_s \times H \times W}.$$9

The final classifier uses global average pooling and a linear layer with sigmoid:

$C_s$0

Training uses binary cross-entropy,

$C_s$1

No additional alignment or contrastive loss is reported; alignment and refinement are handled implicitly by DFM and CBAM.

4. Data regime, optimization, and implementation

The training and evaluation protocol covers three datasets. ISPRS-FD contains 4,742 images spanning ships, vehicles, airports, sea, and land, with image sizes from 256 to 20000 pixels and hybrid forgeries including copy-move, splicing, and diffusion-generated examples. Anti DeepFake, attributed to Zhao et al. 2021, consists of two city domains, Beijing and Seattle, with CycleGAN forgeries. SDGen-Detection, introduced in the paper, uses Potsdam and Vaihingen scenes, the latter in NIR-R-G bands, with forgeries generated by Stable Diffusion 1.4 guided by LLava-1.6-34B prompts. All datasets are randomly split into 50% train and 50% validation/test, and all inputs are resized to $C_s$2 (Qi et al., 25 Jun 2025).

The reported preprocessing uses standard resizing and normalization. Among augmentations, Random Crop is said to help, whereas color, inversion, erase, and mixup degrade performance. A specific result is that mixup reduces F1 by 31.55% on ISPRS-FD. Optimization uses Adam with $C_s$3, initial learning rate $C_s$4, cosine decay, 300 epochs (“cycles”), and batch size 16 per GPU. Training is conducted on a single Nvidia RTX 4090 GPU in OpenMMLab MMPretrain. The experiments include both training from scratch and initialization from several pretraining sources, including ImageNet supervised and self-supervised learning, Million-AID supervised learning, and TOV-RS self-supervised learning.

The paper reports code availability at the GeoX-Lab repository. Parameter count, FLOPs, runtime, and memory are not reported. The description instead characterizes the network as relatively lightweight because it uses a shallow spatial extractor together with moderate convolutional stacks for mapping and refinement.

5. Empirical performance and ablation results

Evaluation uses overall accuracy, precision, recall, and F1, with cross-domain testing for generalization. On ISPRS-FD with a 50% train and 50% test split, SFNet reports OA 81.82, Precision 78.00, Recall 76.17, and F1 76.97. The listed baselines include spatial models such as ViT-Base, Conformer, Inception-v3, and ResNet; frequency models such as FreqNet and FcaNet; and hybrid models such as Geo-DefakeHop and FSLNet. The paper summarizes the gain as 4%–15.18% over prior methods in OA or F1, with especially strong improvement over hybrid baselines (Qi et al., 25 Jun 2025).

Cross-dataset results are more differentiated. On Anti DeepFake with training on Beijing and testing on Seattle, SFNet achieves OA 94.26 and F1 94.01. However, FreqNet reaches approximately 98.3 for OA and F1, and Geo-DefakeHop reaches 100/100 under CycleGAN, while spatial-only models remain much lower, with OA and F1 between roughly 64 and 83. On SDGen-Detection with training on Potsdam and testing on Vaihingen, SFNet attains OA 81.59 and F1 81.32, slightly exceeding FreqNet at OA 80.78 and F1 80.76, while spatial-only models fall into an approximately 66–78 range. The stated takeaway is therefore not that SFNet dominates every setting, but that it maintains balanced robustness across both GAN and diffusion regimes, whereas frequency-only models can be especially strong under CycleGAN.

Ablation studies reinforce the dual-domain interpretation. On ISPRS-FD, the full model yields OA 81.82 and F1 76.97. Removing CBAM drops performance to OA 77.60 and F1 71.82. The image-only variant reports OA 78.57 or 77.10 and F1 71.66 or 71.10 depending on attention; the frequency-only variant reports OA 79.71 or 77.35 and F1 72.47 or 67.49. Among attention mechanisms in the refinement module, CBAM is best, ahead of BAM, SE, and ECA. Another ablation varies the high-pass ratio in HFRI using $C_s$5 for $C_s$6, corresponding to approximately 44.4%, 25.0%, 16.6%, 6.3%, and 0.1%. The reported conclusion is that overly aggressive low-frequency removal, especially above 25%, harms detection, while moderate retention of low- and mid-frequency information is beneficial.

Pretraining studies indicate that self-supervised learning is better than supervised learning on the same domain, which the paper interprets as evidence that image-intrinsic cues matter more than high-level semantics for forgery detection. Remote sensing self-supervised learning with TOV-RS is slightly better than ImageNet self-supervised learning, although ImageNet-SSL still helps the image-domain branch because of its richer semantics.

6. Qualitative interpretation, limitations, and broader usage of the term

The qualitative analyses are centered on Grad-CAM and frequency visualization. Spatial-only models are reported to focus on salient objects regardless of authenticity and can miss subtle artifacts in complex backgrounds. Frequency-only models highlight edge-like responses but show limited diversity in captured patterns, and may miss artifacts in textured terrains or diffusion outputs. SFNet, by fusing both domains, is described as focusing on discriminative regions conditioned on both content and frequency cues. It is reported to detect GAN upsampling signatures and terrain-specific artifacts while filtering natural remote sensing distortions such as shadows and ortho-rectification effects that confound single-cue models (Qi et al., 25 Jun 2025).

The failure cases are also explicit. Extremely subtle diffusion artifacts in homogeneous terrains such as deserts or oceans, as well as heavy real-image distortions, can still confuse the detector. The paper therefore emphasizes careful augmentation and careful frequency masking. Its comparison to earlier spatial-frequency fusion approaches further argues that shallow RGB-frequency concatenation and related fusion methods can suffer from semantic mismatches and scale inconsistencies; SFNet’s stated novelty lies in the spectral-domain frequency extractor with HFRI, HFRF, and FCL, the residual nonlinear DFM alignment mechanism, and CBAM-based hybrid refinement.

The acronym itself is not unique in recent literature. It appears as the spatial-frequency dual-domain translator inside WaveSFNet for spatiotemporal prediction (Cai et al., 24 Mar 2026), as the “Frequency Fusion Network” stage in PRISM for deraining (Xue et al., 30 Sep 2025), and as a 3D MRI classifier for Alzheimer’s disease diagnosis (Yang et al., 22 Jul 2025). Closely related spatio-frequency formulations also appear in SFFNet for remote sensing segmentation (Yang et al., 2024), DSFC-Net for rural road extraction (Zhang et al., 1 Feb 2026), Spatial-Frequency U-Net for diffusion models (Yuan et al., 2023), SFANet for image denoising (Guo et al., 2023), and Vcamba for video camouflaged object detection (Li et al., 31 Jul 2025). This suggests that “SFNet” functions both as a specific model name and as a broader architectural motif: explicit coordination between spatial representations and frequency-aware representations to improve robustness where either domain alone is insufficient.