Combating Pattern and Content Bias: Adversarial Feature Learning for Generalized AI-Generated Image Detection

Haifeng Zhang; Qinghui He; Xiuli Bi; Bo Liu; Chi-Man Pun; Bin Xiao

cs.CVApr 14, 2026

Combating Pattern and Content Bias: Adversarial Feature Learning for Generalized AI-Generated Image Detection

Authors: Haifeng Zhang, Qinghui He, Xiuli Bi, Bo Liu, Chi-Man Pun, Bin Xiao

Organizations: Chongqing Key Laboratory of Image Cognition, School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China · School of Artificial Intelligence, Chongging Key Laboratory of Image Cognition, Chongging University of Posts and Telecommunications, Chongqing 400065, China · Faculty of Science and Technology of the University of Macau, Macau 999078, China · Jinan Inspur Data Technology Co., Ltd., Jinan 250101, China

Abstract

In recent years, the rapid development of generative artificial intelligence technology has significantly lowered the barrier to creating high-quality fake images, posing a serious challenge to information authenticity and credibility. Existing generated image detection methods typically enhance generalization through model architecture or network design. However, their generalization performance remains susceptible to data bias, as the training data may drive models to fit specific generative patterns and content rather than the common features shared by images from different generative models (asymmetric bias learning). To address this issue, we propose a Multi-dimensional Adversarial Feature Learning (MAFL) framework. The framework adopts a pretrained multimodal image encoder as the feature extraction backbone, constructs a real-fake feature learning network, and designs an adversarial bias-learning branch equipped with a multi-dimensional adversarial loss, forming an adversarial training mechanism between authenticity-discriminative feature learning and bias feature learning. By suppressing generation-pattern and content biases, MAFL guides the model to focus on the generative features shared across different generative models, thereby effectively capturing the fundamental differences between real and generated images, enhancing cross-model generalization, and substantially reducing the reliance on large-scale training data. Through extensive experimental validation, our method outperforms existing state-of-the-art approaches by 10.89% in accuracy and 8.57% in Average Precision (AP). Notably, even when trained with only 320 images, it can still achieve over 80% detection accuracy on public datasets.

Figures & tables

Fig. 1: (Color best viewed.) This illustration shows the problems of traditional methods under the asymmetric bias learning phenomenon and our proposed solution. (a) Traditional detectors (e.g., CLIP + classifier) tend to over-rely on the specific generative patterns of particular generative models when performing real/fake classification. (b) Such reliance on specific generative patterns makes the model prone to generalization failure when encountering unseen generative patterns. (c)–(e) MAFL introduces an adversarial mechanism between the real/fake classification task and the specific-pattern classification task, guiding the model to extract common generative features that characterize different generative models. These features are free from model-specific characteristics, thereby enabling stronger generalization when facing unseen generative models.

Fig. 2: t-SNE visualization of features. (a) Visualization of general features extracted by CLIP, labeled by category and authenticity (real/fake). (b) Visualization of features extracted by MAFL, labeled by category and authenticity (real/fake). (c) Visualization of CLIP features mapped to MLP, trained using real and generated images(fake). (d) Visualization of mapped features after Multi-dimensional Adversarial Feature Learning.

Fig. 3: The average frequency spectra and sample examples of images generated by different models are shown. From the frequency domain perspective, we can clearly observe that each model exhibits distinctive generation artifacts. Additionally, at the image level, they display noticeable style differences.

Fig. 4: MAFL Overall Structure. Our framework consists of a feature extractor, a real/generated image classification network, and a bias learning network. Through alternating adversarial training, the bias learning network learns and distinguishes the generative pattern and content differences across various generative models. Meanwhile, the real/generated image classification network guides the feature extractor to learn shared generative features that can “deceive” the bias learning network and remain consistent across different generative models.

Fig. 5: Performance evaluation on the latest unseen generative models.All methods are evaluated under Protocol-1 for a fair comparison. Our method achieves superior detection performance on the latest unseen generative models compared with existing approaches.

Method	Generative Adversarial Networks						Diffusion Models							Others		AP
Method	Pro-GAN	Style-GAN	Cycle-GAN	Star-GAN	Big-GAN	Gau-GAN	SDV1.4	VQDM	SDV1.5	GLIDE	ADM	Midj	Wukong	SAN	Deep-fake	Avg
FreDect [ 18 ] (ICML2020)	100.00	100.00	64.29	99.18	77.16	74.17	51.14	72.08	51.50	40.70	41.58	46.24	50.36	45.75	52.53	64.45
CNNSpot [ 8 ] (CVPR2020)	100.00	100.00	88.81	98.13	77.57	86.39	55.06	65.02	55.52	56.40	68.37	46.82	51.82	59.33	65.65	71.66
Lgrad [ 21 ] (CVPR2023)	99.97	100.00	58.21	99.93	59.15	72.10	51.21	55.28	51.78	66.38	62.43	56.20	48.52	41.38	85.51	67.20
Fusing [ 24 ] (ICIP2022)	100.00	99.99	90.77	99.99	74.14	55.93	76.50	78.63	76.49	58.50	78.67	62.49	72.51	67.33	88.23	78.68
Univfd [ 15 ] (CVPR2023)	100.00	99.97	99.88	99.80	99.23	99.98	53.64	97.87	53.98	78.94	89.42	44.53	64.00	61.41	84.38	81.80

TABLE I: Performance metrics (Average Precision) for various approaches trained using GAN-generated imagery. Other generative frameworks beyond the GANs can be interpreted as the domain for generalization testing.

Method	Generative Adversarial Networks						Diffusion Models							Others		ACC
Method	Pro-GAN	Style-GAN	Cycle-GAN	Star-GAN	Big-GAN	Gau-GAN	SDV1.4	VQDM	SDV1.5	GLIDE	ADM	Midj	Wukong	SAN	Deep-fake	Avg
FreDect [ 18 ] (ICML2020)	100.00	100.00	72.54	83.96	76.05	75.60	50.84	70.89	51.06	40.72	39.66	46.82	50.22	47.26	49.93	63.70
CNNSpot [ 8 ] (CVPR2020)	100.00	100.00	79.25	91.99	71.57	80.56	53.15	60.81	53.83	53.80	62.19	48.03	51.41	55.25	54.56	67.76
Lgrad [ 21 ] (CVPR2023)	99.45	99.86	58.36	86.27	63.05	59.63	54.56	55.47	54.92	55.58	58.53	54.23	51.68	40.63	55.37	63.17
Fusing [ 24 ] (ICIP2022)	99.99	99.98	79.15	99.94	71.45	55.23	68.42	69.03	68.48	58.40	70.54	59.68	66.22	63.47	69.02	73.27
Univfd [ 15 ] (CVPR2023)	99.59	99.94	96.27	84.09	93.32	95.31	51.45	91.34	51.52	69.22	80.35	48.71	57.00	35.47	72.87	76.39

TABLE II: Performance metrics (Accuracy) for various approaches trained using GAN-generated imagery. The statistics referenced in DIRE [ 11 ] are extracted from NPR [ 22 ] , with ”-” representing cases where the approach provided no data for the specific model under consideration.

Method	Generative Adversarial Networks						Diffusion Models							Others		AP
Method	Pro-GAN	Style-GAN	Cycle-GAN	Star-GAN	Big-GAN	Gau-GAN	SDV1.4	VQDM	SDV1.5	GLIDE	ADM	Midj	Wukong	SAN	Deep-fake	Avg
Univfd (CVPR2023)	96.17	77.48	97.69	86.84	85.42	92.38	93.36	98.04	96.81	96.78	92.74	94.30	93.71	92.41	79.22	91.56
NPR (CVPR2024)	88.27	87.46	81.10	100.00	81.27	86.07	79.99	85.65	79.18	71.56	61.58	60.8	83.19	72.30	92.82	80.75
CLIPping (ICMR2024)	92.86	92.24	96.66	80.00	87.53	93.02	98.03	95.74	97.69	98.31	80.09	83.57	95.82	93.97	54.31	89.32
Effort (ICML2025)	97.97	92.11	97.98	97.08	89.17	92.61	99.88	99.97	97.99	94.70	92.30	90.24	97.98	92.72	76.21	93.93
VIB-Net (CVPR2025)	98.66	94.85	98.95	96.67	91.08	94.12	100.00	100.00	98.46	95.95	94.00	92.47	98.91	97.72	71.12	94.86

TABLE III: The AP results of various approaches trained on diffusion-based data are reported. Models not listed under the Diffusion Models (DMs) category are treated as out-of-distribution cases for evaluating cross-framework generalization.

Method	Generative Adversarial Networks						Diffusion Models							Others		ACC
Method	Pro-GAN	Style-GAN	Cycle-GAN	Star-GAN	Big-GAN	Gau-GAN	SDV1.4	VQDM	SDV1.5	GLIDE	ADM	Midj	Wukong	SAN	Deep-fake	Avg
Univfd (CVPR2023)	83.00	64.50	85.50	57.50	71.50	77.50	91.50	93.00	90.00	92.00	86.00	87.50	88.50	82.50	51.00	80.10
NPR (CVPR2024)	74.25	69.17	73.77	98.87	63.85	65.98	85.58	86.87	84.75	75.53	63.31	63.29	85.03	70.09	61.05	74.76
CLIPping (ICMR2024)	92.10	92.01	93.22	87.42	85.48	86.77	97.55	95.23	96.32	96.92	79.80	82.08	93.47	92.66	57.98	88.60
Effort (ICML2025)	95.50	89.00	96.50	82.50	80.50	86.00	97.00	96.50	97.00	90.00	88.10	76.00	97.50	88.50	61.50	88.14
VIB-Net (CVPR2025)	95.00	91.00	97.00	71.00	73.00	87.50	99.00	98.50	97.00	93.50	92.50	80.00	97.50	94.00	49.50	87.73

TABLE IV: The comparison is conducted using classification accuracy, which depends on the chosen decision threshold. To ensure fairness, all methods adopt a unified threshold of 0.5.

CLIP:ViT	AFL	EMLoss	FALoss	LRLoss	Acc	AP
✓					82.17	88.40
✓	✓				85.30	92.31
✓	✓	✓	✓		87.67	94.70
✓	✓	✓		✓	85.40	93.48
✓	✓		✓	✓	87.71	94.59
✓	✓	✓	✓	✓	90.03	96.22

TABLE V: Ablation Study Results

Fig. 6: Test results with different scales of training samples. Our method can still achieve high detection performance when trained with only 300-700 images.

Fig. 7: Robustness evaluation of different methods under perturbations in Protocol-3, with the top row showing the effects of JPEG compression and the bottom row showing the effects of Gaussian blur.

Fig. 8: Training the baseline model (CLIP + classifier) using 2 to 4 datasets. 2: ProGAN + StyleGAN; 3: ProGAN + StyleGAN + CycleGAN; 4: ProGAN + StyleGAN + CycleGAN + GauGAN

Fig. 9: Test results of training MAFL with pairwise combinations of 7 generative models. Training with the same generative model (diagonal) shows poor performance.

Fig. 10: Performance comparison of different MAFL variants. Using CLIP:ViT-L/14 as the backbone achieved the highest performance.

Adversarial Debiasing

See all 8 papers

Reduce the Artifacts Bias for More Generalizable AI-Generated Image Detection
May 14, 2026Yiheng Li, Yang Yang, Zichang Tan +3Synthetic Data Augmentation Mixture of Experts
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China · Sangfor Technologies Inc. +4
FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection
Jul 24, 2026Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1Domain Generalization AI-Generated Image Detection
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia
Diversity Matters: Dataset Diversification and Dual-Branch Network for Generalized AI-Generated Image Detection
Mar 29, 2026Nusrat Tasnim, Kutub Uddin, Khalid MalikAI-Generated Image Detection Cross-Domain Generalization
Korea Aerospace University, 10540, Goyang, Gyeonggi, South Korea · University of Michigan-Flint, 48502, Flint, Michigan, United States

May 14, 2026J/K move · Enter open · S save

Reduce the Artifacts Bias for More Generalizable AI-Generated Image Detection

Yiheng Li, Yang Yang, Zichang Tan, Gao Li+2

School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China · Sangfor Technologies Inc.+4

Synthetic Data Augmentation Mixture of Experts AI-Generated Image Detection

Open paper AskPDF arXiv

As the misuse of AI-generated images grows, generalizable image detection techniques are urgently needed. Recent state-of-the-art (SOTA) methods adopt aligned training datasets to reduce content, size, and format biases, empowering models to capture robust forgery cues. A common strategy is to employ reconstruction techniques, e.g., VAE and DDIM, which show remarkable results in diffusion-based methods. However, such reconstruction-based approaches typically introduce limited and homogeneous artifacts, which cannot fully capture diverse generative patterns, such as GAN-based methods. To complement reconstruction-based fake images with aligned yet diverse artifact patterns, we propose a GAN-based upsampling approach that mimics GAN-generated fake patterns while preserving content, size, and format alignment. This naturally results in two aligned but distinct types of fake images. However, due to the domain shift between reconstruction-based and upsampling-based fake images, direct mixed training causes suboptimal results, where one domain disrupts feature learning of the other. Accordingly, we propose a Separate Expert Fusion (SEF) framework to extract complementary artifact information and reduce inter-domain interference. We first train domain-specific experts via LoRA adaptation on a frozen foundational model, then conduct decoupled fusion with a gating network to adaptively combine expert features while retaining their specialized knowledge. Rather than merely benefiting GAN-generated image detection, this design introduces diverse and complementary artifact patterns that enable SEF to learn a more robust decision boundary and improve generalization across broader generative methods. Extensive experiments demonstrate that our method yields strong results across 13 diverse benchmarks. Codes are released at: https://github.com/liyih/SEF_AIGC_detection.

Abstract

Adversarial Debiasing

Explore similar work

Reduce the Artifacts Bias for More Generalizable AI-Generated Image Detection

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection

Diversity Matters: Dataset Diversification and Dual-Branch Network for Generalized AI-Generated Image Detection