Multi-Domain Perception Transformer for Generalized Forgery Image Detection

Abstract

With the rapid advancement of generative AI (AIGC) technology, synthetic images are increasingly approaching real pictures in terms of resolution and semantic consistency. Traditional detection methods face numerous challenges, such as insufficient cross-modal generalization capabilities and difficulty in identifying hidden generative traces. Existing solutions primarily design feature extractors for single generative models, struggling to address the complexity of multimodal forgeries. Therefore, we propose a multi-domain feature fusion Transformer network that integrates spatial, frequency, and wavelet transform features and introduce a cross-domain feature fusion module (CDAF) to detect subtle forgery traces in deepfake images. This model demonstrates superior detection performance on current forged images generated by generative adversarial networks (GANs) and diffusion models while exhibiting enhanced robustness.

1. Introduction

With the rapid development of deep generative technologies—such as Generative Adversarial Networks (GANs) [1], Variational Autoencoders (VAEs) [2], and Diffusion Models (DMs) [3,4]—it has become increasingly challenging to distinguish high-fidelity deepfake images from authentic ones. Deepfake technology has proliferated rapidly, giving rise to serious security concerns, including misuse on social media, fabrication of personal information, and political fraud. These threats pose severe challenges to public information security, personal privacy, and societal trust. Consequently, there is an urgent need for efficient and robust deepfake detection techniques.

In the early stages of deepfake generation, due to technological limitations, synthetic images often exhibited obvious visual artifacts—such as visible splicing boundaries, inconsistent facial structures, abnormal color distributions, and other generation-related anomalies. As a result, many early detection approaches adopted standard binary classification models or simple convolutional neural networks (CNNs) to distinguish real from fake images. For example, Tyagi et al. [5] proposed MiniNet, a lightweight fully convolutional neural network that emphasizes image-level features for detecting forgeries such as copy-move and image splicing. Gaurav et al. [6] employed a CNN combined with error-level analysis to identify image tampering by evaluating the effects of resaving images at specific JPEG quality levels and quantifying discrepancies in compression artifacts. Muthukumar et al. [7] utilized the DenseNet121 architecture with hyperparameter optimization to adaptively extract shared discriminative features for detecting forged face images. Yang et al. [8] enhanced texture artifacts induced by post-processing by applying a guided filter with a saliency map as the guidance image, thereby amplifying texture differences between real and fake samples to improve discriminability. Wang et al. [9] introduced semantic masks to direct the detector’s attention toward key facial regions, enabling their CNN-based model to achieve improved detection performance on forged face images. However, these methods typically perform well only on data distributions similar to their training sets and often lack generalization capabilities. Moreover, as deepfake generation techniques have advanced—particularly with the iterative improvements in GANs and the emergence of diffusion models—the resulting synthetic images now exhibit richer details, more coherent structures, and fewer detectable artifacts. This evolution renders traditional CNN-based detection approaches increasingly ineffective and insufficient.

To address these limitations, several recent works have explored more robust and generalizable strategies. Frank et al. [10] leveraged GANs to generate synthetic images and introduced perturbations in the discrete cosine transform (DCT) domain to suppress frequency-level artifacts in real images. Their classifier was trained on both original and perturbed images, demonstrating enhanced generalization across unseen forgery types. Gu et al. [11] proposed a progressive reinforcement learning framework that jointly exploits RGB and frequency-domain cues by decomposing images into frequency-specific components, thereby isolating genuine and forged traces in the spectral domain. However, frequency-based methods often exhibit limited robustness against common post-processing operations—such as upsampling, downsampling, or additive noise—which can significantly degrade their discriminative power. Recognizing that over-reliance on frequency artifacts may hinder generalization, Jeong et al. [12] argued that ignoring such artifacts during training can improve a detector’s adaptability across diverse GAN architectures. They proposed FrePGAN, a framework that simultaneously trains a perturbation generator and a deep forgery detector, substantially improving cross-model generalization. Shiohara [13] introduced a self-blended image synthesis technique to train detectors to identify boundary inconsistencies between source and blended regions in forged faces. Shi et al. [14] adopted a reconstruction-based approach: they first pre-trained models on real facial data using masked image modeling, then applied these models to reconstruct fake images and classified samples based on the discrepancy between the input and reconstructed outputs. While promising, such reconstruction-based methods often require intermediate resampling steps, increasing computational complexity. Moreover, their detection performance heavily depends on the quality and relevance of the reconstructed features; poorly chosen reconstruction objectives can even degrade detection accuracy.

Continuous iterations of Generative Adversarial Networks (GANs) and diffusion models have produced synthetic images that are increasingly indistinguishable from real ones. Detecting deepfakes remains challenging due to the diversity of forgery techniques, the high visual fidelity of generated content, and the limited cross-domain generalization of existing detection models. Despite significant advances in the field, accurately distinguishing authentic images from deepfakes remains a formidable task. To address these challenges, we propose a multi-domain Transformer network that integrates features from the spatial domain, the frequency domain, and the wavelet transform for deepfake image detection. Our method first extracts spatial representations, frequency-domain spectrograms, and wavelet coefficients from the input image. These heterogeneous features are then projected into a unified embedding space through a carefully designed feature embedding module. Subsequently, a multi-branch Transformer encoder processes each domain-specific feature stream in parallel, while cross-domain attention mechanisms facilitate interaction and fusion across domains. Finally, the fused high-level semantic features are fed into a classifier to discriminate between genuine and deepfake images. The proposed model demonstrates enhanced robustness against state-of-the-art fake images generated by both GANs and diffusion models and exhibits superior cross-generator generalization capabilities.

The specific contributions of this work are as follows:

• We propose a multi-domain feature fusion Transformer network that jointly leverages spatial, frequency, and wavelet-domain representations—capturing multi-scale characteristics through wavelet transforms—to uncover subtle and otherwise imperceptible forgery traces in deepfake images.
• Features are extracted using EfficientNet independently from the spatial domain, the frequency domain, and the wavelet domain. These features are then fused via a novel Cross-Domain Attention Fusion (CDAF) module, followed by a Swin Transformer that effectively identifies fine-grained forgery artifacts.
• We conduct comprehensive evaluations across multiple benchmark datasets for AI-generated image detection, covering a wide spectrum of state-of-the-art generative models. Extensive experimental results demonstrate that our approach significantly outperforms existing state-of-the-art detectors in both in-domain and cross-domain settings.

2. Related Works

Early work in deepfake detection primarily focused on identifying specific physiological or physical inconsistencies. Li et al. [15] exposed anomalies in fake videos by detecting abnormal blinking patterns. Similarly, Yang et al. [16] leveraged inconsistencies in head pose to distinguish real from fabricated footage. With the advancement of deep learning, research shifted toward designing more robust network architectures capable of learning universal forgery patterns. Rossler et al. [17] released the large-scale benchmark dataset FaceForensics++ and systematically evaluated the performance of various CNN architectures—including Xception [18] and ResNet [19]—on deepfake detection tasks. Their study revealed that the Xception network exhibited exceptional generalization capabilities for this task. Nguyen et al. [20] (2021) proposed Capsule-Forensics, which employs capsule networks to capture spatial hierarchies among local features, thereby identifying geometric distortions introduced during the forgery process. Chen et al. [21] introduced Self-Supervised Deepfake Detection (SSDD), which constructs positive–negative sample pairs (real–real vs. real–fake) to learn forgery-invariant representations. Although these methods achieve strong performance on specific datasets, their effectiveness significantly degrades when faced with unseen generative models or cross-dataset scenarios, revealing limitations in generalization.

Consequently, recent research has focused on improving the robustness and universality of detection models. Ren et al. [22] proposed exploiting GAN fingerprints—unique, learnable artifacts left by different generative models in their outputs—that persist even after post-processing. Separately, researchers have found that incorporating frequency-domain analysis can reveal subtle generation artifacts by examining the spectral characteristics of images. Genuine and forged images exhibit statistically significant differences in the frequency domain, particularly in high-frequency components. For instance, Zhao et al. [23] identified anomalies in high-frequency coefficients (e.g., Discrete Cosine Transform coefficients) within deepfake images and proposed the Frequency-aware Deepfake Detection (F3-Net) framework, which substantially improves cross-domain detection performance. Kim et al. [24] introduced FReTA, a detection method based on representation learning and knowledge distillation. It leverages pre-trained models to rapidly adapt to performance-based deepfakes via transfer learning. Aneja et al. [25] proposed a few-shot learning approach grounded in transfer learning, framing deepfake detection as a domain-shift problem: an encoder is first trained on a source domain and then fine-tuned using only a few samples from the target domain. During inference, classification is performed based on the distance between the latent representations of real and fake samples. Wang et al. [26] proposed the RFM framework, which enhances the training efficacy of CNNs through attention mechanisms. Their method first generates a forgery attention map for the backbone network using the original image, then performs data augmentation by erasing suspicious regions before retraining the model.

In recent years, the widespread adoption of Transformers [27] in natural language processing and computer vision—owing to their performance surpassing that of CNNs—has inspired their application in deepfake detection. Wodajo et al. [28] proposed CvT, which combines CNNs and Vision Transformers (ViTs) as dual backbones for feature extraction to authenticate facial images. Heo et al. [29] integrated knowledge distillation into a detection model with a Vision Transformer as the backbone, using the CNN-based EfficientNet as the teacher model. Park et al. [30] explored a novel model based on the Swin-Transformer architecture to distinguish between real and fake images generated by convolutional neural networks (CNNs) and generative adversarial networks (GANs), preventing the misuse of AI images through procedural discrimination/legal sanctions. Xu et al. [31] proposed an efficient visual Transformer model that combines CNN features derived from vector concatenation with a patch-based localization method to identify fake regions by interacting with all locations, for DeepFake detection to extract local and global features. In this paper, we propose an efficient forgery detection network that adopts a Transformer as the backbone architecture. Specifically, we employ a Vision Transformer (ViT) to extract global semantic features and integrate an attention mechanism to localize forged regions. Furthermore, our method fuses multimodal features from both the spatial domain and the frequency domain to effectively detect images generated by diverse classes of generative models.

3. Materials and Methods

This section mainly introduces our proposed model and its details.

3.1. Overview

At present, deepfake images are predominantly generated by more stable GAN models and advanced diffusion models, resulting in richer and more accurate fine-grained details. To accurately detect these increasingly indistinguishable deepfakes, we propose a multi-domain Transformer network framework that integrates spatial, frequency, and wavelet transform features. This framework jointly utilizes representations from the image’s spatial, frequency, and wavelet domains. Deepfakes inevitably exhibit multi-domain inconsistencies, such as spatial artifacts, frequency anomalies, and multi-scale distortions. A carefully designed embedding module maps these features into a unified feature space. Subsequently, a multi-branch Transformer encoder processes features from each domain while enabling interaction and fusion through cross-domain attention mechanisms. Finally, the fused high-level semantic features distinguish between real and fake images. The model framework is illustrated in Figure 1.

3.2. Multi-Domain Feature Extraction

In the spatial domain extraction section, a lightweight EfficientNet-B0 network is employed for feature extraction, capturing inconsistent feature information such as image texture, edges, and local features. To enhance edge sensitivity, a gradient-based map (Laplacian map) is concatenated to the RGB channels before convolution, capturing fine-grained local distortions commonly seen in blending or diffusion-based artifact boundaries.

Spatial domain features:

$$F_{\mathrm{s} }\in {\mathbb{R}}^{B\times C_{sp}\times H\times W}$$

(1)

where $B$ is the batch size, $C_{sp}$ represents the spatial feature channel, and $H\times W$ is the spatial resolution of the image.

In the frequency domain extraction stage, most image generation models—such as GANs and diffusion models—tend to produce smoother, more natural low-frequency content. However, they often struggle with high-frequency details like textures and edge sharpness, leading to anomalies such as excessive smoothing and periodic artifacts. Here, we compute a block-wise Discrete Cosine Transform (DCT), retaining only mid-frequency and high-frequency coefficients to eliminate semantic redundancy.

Frequency domain features:

$$\begin{gathered} F_f=f_f\left(F_{freq}\right) \\ {F}_{freq}(u,v)=\sum _{x=0}^{N-1} \sum _{y=0}^{N-1} I(x,y)\mathrm{c}\mathrm{o}\mathrm{s}{\displaystyle \frac{(2x+1)u\pi }{2N}}\mathrm{c}\mathrm{o}\mathrm{s}{\displaystyle \frac{(2y+1)v\pi }{2N}} \end{gathered}$$

(2)

where $u,v$ are frequency domain coordinates. The frequency domain coefficients obtained by performing a two-dimensional discrete cosine transform on the $F_{freq}(u,v)$ image I constitute the frequency domain feature map.

The wavelet domain feature extraction component leverages the multi-scale and multi-directional properties of wavelet transforms to effectively capture local image details, edge structures, and texture information. captures low-frequency and high-frequency subband (LL, LH, HL, HH) feature information F_wave through multi-level discrete wavelet transforms (DWT). Here, we employ EfficientNet for feature detail extraction, the same model used for extracting spatial and frequency domain features. Frequency domain features:

$$F_w=f_f\left(F_{wave}\right)$$

(3)

3.3. Cross-Domain Attention Feature Fusion Module

The spatial domain features $F_{\mathrm{s} }$, frequency domain features $F_f$, and wavelet domain features $F_w$ obtained through lightweight EfficientNet convolutional computations require feature fusion before proceeding to the next step. Here, we employ a cross-domain attention fusion model for feature integration, as illustrated in Figure 2. The cross-domain attention fusion is defined as:

$$\mathrm{C}\mathrm{D}\mathrm{A}\mathrm{F}(Q,K,V)=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\displaystyle \frac{Q{K}^T}{\sqrt{d}}}\right)V$$

(4)

where $Q$ represents the spatial domain feature $F_{\mathrm{s} }$, $K$ is the concatenated feature ($F_f$, $F_w$) from the frequency domain and wavelet domain, and $V=K$.

The final result is the fused feature $F_{\mathrm{fused} }$. It is defined as:

$$F_{\mathrm{fused} }=\sigma \left(W_s\right)F_s+\sigma \left(W_f\right)F_f+\sigma \left(W_w\right)F_w$$

(5)

where is $\sigma (\cdot )$ a sigmoid gate.

The fused representation is then fed into the Swin Transformer for hierarchical global reasoning.

3.4. Transformer Module

In the configuration of the backbone feature detection network, to better uncover subtle artifacts in forged images, we employ the Swin Transformer as the backbone network for detecting genuine and forged features after feature fusion, as shown in Figure 3. Swin Transformer is built by replacing the standard multi-head self-attention (MSA) module in a Transformer block with a module based on shifted windows, with other layers kept the same. A Swin Transformer block consists of a shifted window-based MSA module, followed by a 2-layer MLP with GELU nonlinearity in between. A LayerNorm (LN) layer is applied before each MSA module and each MLP, and a residual connection is applied after each module.

Overall loss function:

$${\mathcal{L}}_{total}={\lambda }_1{\mathcal{L}}_{spatial}+{\lambda }_2{\mathcal{L}}_{freq}+{\lambda }_3{\mathcal{L}}_{wavelet}+{\lambda }_4{\mathcal{L}}_{fusion}+{\lambda }_5{\mathcal{L}}_{cls}$$

$${\mathcal{L}}_{spatial}=\mathrm{CrossEntropy}(y_{spatial},y_{true})$$

$${\mathcal{L}}_{freq}=\mathrm{CrossEntropy}(y_{freq},y_{true})$$

$${\mathcal{L}}_{wavelet}=\mathrm{CrossEntropy}(y_{wavelet},y_{true})$$

$${\mathcal{L}}_{fusion}=D_{KL}(A_{fusion}\parallel \mathrm{U}\mathrm{n}\mathrm{i}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m})$$

$${\mathcal{L}}_{cls}=-(1-ϵ)y\mathrm{l}\mathrm{o}\mathrm{g}\left(p\right)-ϵ(1-y)\mathrm{l}\mathrm{o}\mathrm{g}(1-p)$$

where $y_{spatial}$ represents the spatial domain feature, $y_{freq}$ represents the frequency domain feature, $y_{wavelet}$ represents the wavelet transform feature, $y_{true}$ represents the true label, and $A_{fusion}$ represents the attention weight vector generated in the fusion module.

4. Experiments and Results

4.1. Dataset

In terms of dataset selection, to better verify the reliability and robustness of the model and consider the impact of different types of generative models on the detection model, this study used multiple datasets for verification experiments, including the common face datasets DFDC [32] and Celeb-DF [33], as well as ForenSynths [34], based on various types of images. All images were resized to 256 × 256 pixels. In splitting the dataset, we used 80% for model training and 20% for testing.

Celeb-DF is a large-scale challenging dataset for deep fake forensics. The dataset includes 590 original videos collected from YouTube and 5639 fake videos generated by DeepFake, with people of different ages, races, and genders in the videos.

DFDC, the Deepfake Detection Challenge dataset, is a large dataset released by Meta to measure the progress of Deepfake detection technology. This dataset is a deep face detection dataset consisting of more than 100,000 fake videos created from 19,154 real videos and fully considers the diversity of subjects and backgrounds in real scenes (skin color, gender, lighting conditions, etc.).

ForenSynths: This dataset contains fake images generated by 11 different convolutional neural network (CNN)-based image generator models. These models cover commonly used architectures today (ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, etc.).

On the other hand, we also considered and selected some fake image data generated based on diffusion models, including PNDM [35], Guided [36], DALL-E [37], and VQ-Diffusion [38].

4.2. Evaluation Metrics and Implementation Details

In the initial spatial, we employ three parallel EfficientNet-B3 encoders to extract features from the spatial, frequency, and wavelet domain representations of the input image, respectively. Each EfficientNet-B0 is initialized with ImageNet-pretrained weights and fine-tuned end-to-end. The output feature maps from each branch have a spatial resolution of H/32 × W/32 and a channel dimension of 384 (i.e., the final stage output before global pooling in EfficientNet-B3). These features are then projected via lightweight 1 × 1 convolutions into a unified embedding space of dimension d = 256, enabling cross-domain alignment. The three domain-specific embeddings are concatenated along the channel dimension and fed into a modified Swin Transformer backbone (Tiny configuration: window size = 8, 4 stages, depths = [2, 2, 6, 2], embed dim = 96). To accommodate the fused input, the first linear patch embedding layer of Swin Transformer is reconfigured to accept an input dimension of 3 × 256 = 768, and its weights are initialized using Xavier uniform initialization. Subsequent transformer blocks maintain the standard Swin architecture, ultimately producing a high-resolution feature map that is upsampled via a transposed convolution head for output forgery localization. EfficientNet employs the Adam optimizer with a learning rate of 1 × 10⁻⁴. The backbone Swin Transformer network utilizes the Adamw optimizer with a learning rate of 5 × 10⁻⁵. Here, the spatial loss ${\lambda }_1$, frequency domain loss ${\lambda }_2$, and wavelet domain loss ${\lambda }_3$ are all set to 0.3, the fusion loss ${\lambda }_4$ is 0.2, and the backbone transformer network loss ${\lambda }_5$ is 1.0. The batch size is set to 16, and the number of epochs is 100. We evaluate the proposed method using Average Precision (AP) and Accuracy (ACC) as assessment metrics. All experiments were trained on a server equipped with dual NVIDIA RTX 3090Ti GPUs, AMD Threadripper 2950X CPU, and 64 G RAM.

4.3. Comparisons with State-of-the-Art Methods

GAN-generated fake images commonly exhibit certain telltale signs or anomalous features, as shown in Figure 4. Their fabricated texture characteristics are relatively conspicuous, making them easier for detection models to identify. However, fake images generated using the Stable Diffusion model, as illustrated in Figure 5, possess forgery traces that are difficult to discern at a glance. Consequently, models that extract only a single feature from an image struggle to accurately detect the subtle, deep-seated forgery traces within them.

To uncover hidden traces in deepfake images, we performed frequency domain visualization analysis on the forged images, as shown in Figure 6. In the image frequency domain visualization spectrum, the forgery features of the forged image are magnified, and its spectral features are clearly different from those of the real image.

In the model performance comparison test, we first used the CFDC and celeb-DF datasets based on face images to analyze the model performance. Simultaneously, we tested the impact of different modules on the overall network framework. As shown in

Table 1. Comparing with advanced models on facial forgery datasets.

Models	Celeb-DF		DFDC
	ACC	AP	ACC	AP
Xception	67.4	65.8	66.3	68.3
F3-Net	77.2	76.7	75.7	76.0
TAN-GFD [39]	87.7	90.3	84.3	85.8
WMamba [40]	92.3	91.4	90.5	90.0
Ours(S+F)	92.7	91.8	91.7	91.1
Ours(S+F+W)	95.1	94.3	93.9	93.0
Ours(S+F+W+CDAF)	96.3	95.9	95.4	94.2

, our model performed better in both accuracy and average precision. The addition of wavelet domain features significantly improved detection performance through multi-scale feature extraction, and the inclusion of CDAF further enhanced the ability to identify image forgery traces.

In model testing experiments for detecting deepfake images generated by GANs, we employed the Forensynths dataset to evaluate images produced by various GAN models, as shown in

Table 2. Performance comparison with other state-of-the-art models on the ForenSynths dataset based on accuracy (ACC)/average precision (AP).

Methods	ProGAN	StyleGAN	StyleGAN2	BigGAN	CycleGAN	StarGAN	GauGAN	Deepfake	Mean
Wang	64.6/92.7	52.8/80.8	75.7/96.3	50.7/70.2	58.1/79.3	51.2/81.7	53.6/84.7	50.3/51.5	57.1/79.7
Fank	85.7/81.3	73.1/68.5	75.0/70.9	76.9/70.8	86.5/80.8	85.0/77.0	67.3/65.3	50.1/55.3	75.0/71.2
F3-Net	87.8/82.4	80.3/84.7	82.2/87.9	65.5/73.4	81.2/89.7	87.8/90.4	57.0/59.5	59.9/83.0	75.2/81.4
BiHPF	87.4/89.3	71.5/74.1	77.0/81.1	82.6/80.6	86.0/86.6	93.8/95.5	75.3/84.7	53.5/55.8	78.4/81.0
FrePGAN	95.3/97.1	82.0/90.9	72.2/93.8	66.7/69.4	69.7/71.1	97.3/99.0	53.7/55.0	62.7/80.1	75.0/82.1
UniFD	98.3/99.8	78.5/92.8	75.4/96.0	89.1/94.7	91.9/98.0	96.1/99.3	92.6/98.3	80.8/90.2	88.1/96.1
FreqNet	99.2/99.9	90.4/98.0	85.8/98.3	89.7/96.4	96.7/99.1	97.5/99.4	88.3/98.9	81.9/92.7	91.2/98.0
Ours	99.6/99.9	95.7/98.5	91.7/99.1	95.3/98.9	98.5/99.7	98.8/99.6	93.8/99.4	92.3/97.9	96.0/99.1

. Compared to other state-of-the-art models, our model demonstrates greater robustness.

Compared to GAN-based deepfakes, diffusion model-generated deepfakes are more difficult to detect. They exhibit more natural edge transitions and more concealed forgery traces. Most detection models perform poorly against these challenges, as shown in

Table 3. Accuracy and average precision comparisons with state-of-the-art methods on the diffusion model dataset.

Methods	PNDM	Guided	DALL-E	VQ-Diffusion	Mean
Wang	50.8/90.3	54.9/66.6	51.8/61.3	50.0/71.0	51.8/72.3
Fank	44.0/38.2	53.4/52.7	57.1/62.8	52.0/66.3	51.6/55.0
F3-Net	72.8/80.5	69.7/72.1	72.3/80.0	91.8/94.7	76.7/81.8
UniFD	75.3/92.5	75.7/85.1	89.5/96.8	83.5/97.7	81.0/93.0
FreqNet	89.3/97.0	81.2/92.0	94.8/98.3	92.0/97.3	89.3/96.2
Ours	94.1/96.3	82.0/91.7	97.3/99.2	95.4/99.1	92.2/96.6

. Our proposed multi-domain fusion Transformer network simultaneously searches for forgery traces in the spatial, frequency, and wavelet domains of the manipulated images, significantly enhancing the model’s detection capabilities. Even when compared to other excellent models, our approach consistently demonstrates outstanding performance.

5. Discussion and Conclusions

In deepfake image detection tasks, most existing research has focused on verifying the authenticity of individual facial images, with only a limited body of work addressing general-purpose fake detection. Among these, the majority target the authentication of images synthesized by Generative Adversarial Networks (GANs). However, such methods often exhibit strong performance only on images generated by specific architectures or training paradigms, and their effectiveness significantly degrades when applied to images produced by more advanced generators—particularly stable GAN variants and, more recently, diffusion models. To address these limitations, this paper proposes a multi-domain Transformer network that jointly leverages spatial, frequency, and wavelet-domain representations for robust detection of deepfakes. Specifically, our method first extracts spatial features, frequency-domain spectrograms (obtained via the Fourier transform), and multi-scale wavelet coefficients from the input image. These heterogeneous signals are then projected into a unified latent space through a dedicated embedding module. A multi-branch Transformer encoder processes each domain separately, while cross-domain attention mechanisms facilitate dynamic interaction and fusion of complementary cues across domains. The resulting fused high-level features are subsequently used for binary classification between real and deepfake images.

Extensive experiments demonstrate that our approach achieves superior robustness against a wide spectrum of generative models—including both GANs and state-of-the-art diffusion-based synthesizers—and exhibits markedly improved cross-generator generalization compared to existing methods. This suggests that integrating multi-domain signatures can effectively capture intrinsic artifacts left by diverse generative processes, which are often imperceptible in the spatial domain alone. We hope this work will inspire further exploration of hybrid signal representations and innovative architecture designs for deepfake detection. Future research directions will focus on three key areas: (1) model miniaturization to enable deployment on resource-constrained devices; (2) optimization for low-power hardware platforms to facilitate real-time detection in edge computing environments; and (3) extension to handle compressed and re-encoded media commonly encountered on social platforms, with ultimate progression toward comprehensive video-level deepfake forensics. These advancements will collectively contribute to building a more robust, efficient, and holistic deepfake detection ecosystem capable of addressing real-world deployment challenges.