Organizations: Univ. Polytechnique Hauts-de-France, Valenciennes, France · Institute of Applied Sciences and Intelligent Systems – Lecce, Italy · Department of Innovation Engineering, University of Salento, Italy · Sorbonne Center for Artificial Intelligence, Sorbonne University Abu Dhabi
Detecting AI-generated images remains a significant challenge because detectors trained on specific generators often fail to generalize to unseen models; however, while pixel-level artifacts vary across models, frequency-domain signatures exhibit greater consistency, providing a promising foundation for cross-generator detection. To address this, we propose SPARK-IL, a retrieval-augmented framework that combines dual-path spectral analysis with incremental learning by utilizing a partially frozen ViT-L/14 encoder for semantic representations alongside a parallel path for raw RGB pixel embeddings. Both paths undergo multi-band Fourier decomposition into four frequency bands, which are individually processed by Kolmogorov-Arnold Networks (KAN) with mixture-of-experts for band-specific transformations before the resulting spectral embeddings are fused via cross-attention with residual connections. During inference, this fused embedding retrieves the k nearest labeled signatures from a Milvus database using cosine similarity to facilitate predictions via majority voting, while an incremental learning strategy expands the database and employs elastic weight consolidation to preserve previously learned transformations. Evaluated on the UniversalFakeDetect benchmark across 19 generative models -- including GANs, face-swapping, and diffusion methods -- SPARK-IL achieves a 94.6% mean accuracy, with the code to be publicly released at https://github.com/HessenUPHF/SPARK-IL.
Figures & tables
Figure 1: Comparison of detection accuracy and model efficiency on the UniversalFakeDetect benchmark. (a) Efficiency comparison in terms of number of parameters. (b) Mean accuracy across 19 generative models.
Figure 2: SPARK-IL architecture. Dual spectral paths process ViT features and RGB pixels through multi-band KAN-FFT blocks. Cross-attention fuses the spectral embeddings, which are used for classification during training and stored in Milvus for retrieval-based inference.
Figure 3: Multi-band KAN-FFT block. Input features are transformed via FFT, converted to log-magnitude spectrum, partitioned into four frequency bands, processed through dedicated KAN layers, and fused into a spectral embedding.
Methods
Ref
GAN
DeepFakes
Low level
Perceptual loss
LDM
GLIDE
DALL·E
mAcc
Pro- GAN
Cycle- GAN
Big- GAN
Style- GAN
Gau- GAN
Star- GAN
SITD
SAN
CRN
IMLE
200 steps
200 w/cfg
100 steps
100 27
50 27
100 10
Freq-spec
WIFS2019
49.90
99.90
50.50
49.90
50.30
99.70
50.10
50.00
48.00
50.60
50.10
50.90
50.40
50.40
50.30
51.70
51.40
50.40
50.00
55.45
Co-occurrence
Electr. Img
97.70
97.70
53.75
92.50
51.10
54.70
57.10
63.06
55.85
65.65
65.80
60.50
70.70
70.55
71.00
70.25
69.60
69.90
67.55
66.86
CNN-Spot
CVPR2020
99.99
85.20
70.20
85.70
78.95
91.70
53.47
66.67
48.69
86.31
86.26
60.07
54.03
54.96
54.14
60.78
63.80
65.66
55.58
69.58
Patchfor
ECCV2020
75.03
68.97
68.47
79.16
64.23
63.94
75.54
75.14
75.28
72.33
55.30
67.41
76.50
76.10
75.77
74.81
73.28
68.52
67.91
71.24
F3Net
ECCV2020
99.38
76.38
65.33
92.56
58.10
100.00
63.48
54.17
47.26
51.47
51.47
69.20
68.15
75.35
68.80
81.65
83.25
83.05
66.30
71.33
Table 1: Comparison across GAN, DeepFakes, low-level, perceptual-loss, and diffusion-based generators.
Figure 4: t-SNE visualization of SPARK-IL RAG embedding distributions for real (green) and fake (red) samples across different generative models.
Nb shots
GAN
DeepFakes
Low level
Perceptual
LDM
GLIDE
DALL·E
mAcc
ProGAN
CycleGAN
BigGAN
StyleGAN
GauGAN
StarGAN
SITD
SAN
CRN
IMLE
200
200
100
100
50
100
1
99.30
96.90
98.60
95.90
98.80
98.40
93.20
94.90
63.90
95.10
94.90
68.40
98.90
95.50
98.90
94.60
95.10
95.40
98.20
92.80
3
99.55
97.20
98.90
96.20
99.10
98.70
93.70
95.30
64.50
95.60
95.40
69.00
99.10
95.80
99.20
94.90
95.40
95.70
98.50
93.60
5
99.70
97.40
99.05
96.45
99.25
98.90
94.00
95.60
64.90
95.90
95.70
69.30
99.25
95.95
99.40
95.10
95.60
95.90
98.70
94.10
10
99.85
97.55
99.15
96.65
99.35
99.00
94.15
95.70
65.10
96.00
95.85
69.55
99.35
96.05
99.50
95.25
95.75
96.05
98.80
94.40
15
99.90
97.58
99.18
96.75
99.38
99.05
94.25
95.75
65.15
96.05
95.95
69.70
99.38
96.08
99.55
95.35
95.85
96.15
98.85
94.50
Table 2: Effect of the number of retrieved neighbors (Nb shots) on SPARK-IL performance across all generator families on the UniversalFakeDetect benchmark. Performance consistently improves with more retrieved samples and stabilizes beyond 10–15 shots.
Method
Backbone
Params (M)
GFLOPs
REVEAL
ViT-L/14 (partial)
303.0
45.00
C2P-CLIP
ViT-L/14 (partial)
304.0
46.00
RINE
ViT-L/14
434.0
52.00
FatFormer
ViT-L/14
493.0
58.00
SPARK-IL (Ours)
ViT-L/14 (partial)
315.0
60.33
Table 3: Comparison of computational complexity among recent methods, reporting the number of parameters and GFLOPs for a 224 × 224 input.
Model Variant
Pixel FFT
Feature FFT
Retrieval + Voting
KAN
mAcc (%)
(A) ViT Baseline
–
–
–
–
81.5
(B) + Pixel Spectral
✓
–
–
–
85.2
(C) + Dual Spectral
✓
✓
–
–
88.7
(D) + MLP Classifier
✓
✓
–
–
89.1
(E) + KAN (w/o Retrieval)
✓
✓
–
✓
90.3
(F) SPARK-IL (Full)
✓
✓
✓
✓
94.6
Table 4: Ablation study on the UniversalFakeDetect benchmark. The baseline corresponds to a standard Vision Transformer fine-tuned for synthetic image detection. Each row incrementally adds spectral modeling, classifier design, and retrieval-based reasoning.
Generalization remains a critical bottleneck in AI-generated image detection. Because many modern generators are proprietary or adversarially modified, existing detectors overfit to the low-level textural patterns of accessible training data, resulting in severe failures on unseen domains. Conventional regularization techniques (e.g., L1/L2 norms, Dropout) apply indiscriminate parametric constraints and fail to provide the domain-invariant structure necessary for cross-generator robustness. To address this, we propose Feature-Augmented Implicit Regularization (FAIR). FAIR introduces an orthogonal, macro-structural prior, specifically, Scene Composition Structure (SCS), during training to geometrically constrain the model's optimization trajectory. By augmenting the primary feature space with domain-invariant SCS features, FAIR explicitly penalizes texture-biased shortcut learning. Crucially, this structural prior is entirely discarded at inference, yielding a smoothed, generalized decision boundary with zero architectural or computational overhead. Extensive evaluations across five massive benchmarks demonstrate that integrating FAIR into state-of-the-art detectors significantly improves cross-generator generalization, boosting accuracy by up to 8.04% and establishing new state-of-the-art robustness in zero-shot transfer scenarios.
Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia
The rapid advancement of generative models has made the detection of AI-generated images a critical challenge for both research and society. Recent works have shown that most state-of-the-art fake image detection methods overfit to their training data and catastrophically fail when evaluated on curated hard test sets with strong distribution shifts. In this work, we argue that it is more principled to learn a tight decision boundary around the real image distribution and treat the fake category as a sink class. To this end, we propose SimLBR, a simple and efficient framework for fake image detection using Latent Blending Regularization (LBR). Our method significantly improves cross-generator generalization, achieving up to +24.85% accuracy and +69.62% recall on the challenging Chameleon benchmark. SimLBR is also highly efficient, training orders of magnitude faster than existing approaches. Furthermore, we emphasize the need for reliability-oriented evaluation in fake image detection, introducing risk-adjusted metrics and worst-case estimates to better assess model robustness. All code and models will be released on HuggingFace and GitHub.
Aayush Dhakal, Subash Khanal, Srikumar Sastry +4
Washington University in St. Louis · Oak Ridge National Laboratory
The rapid advancement of generative models has blurred the boundary between synthetic and real imagery, creating an urgent need for reliable deepfake detection. Yet most existing approaches rely on massive real--fake datasets, which are increasingly difficult to maintain as new generators continue to emerge. In this work, we investigate how much information about image authenticity is already encoded in modern multimodal vision representations. We find that frozen multimodal encoders naturally separate real and synthetic images in their embedding space, enabling a simple linear classifier to achieve strong performance without task-specific fine-tuning. Motivated by this observation, we develop a representation-aware data curation strategy that selects a compact set of representative generators for training. The resulting training set contains only 10K images, compared to 288K in AIGIBench and 4M in OpenFake, while improving robustness to unseen generators and distribution shifts. We additionally introduce RealWorldBench, a benchmark consisting of modern camera photographs, contemporary stock images, and outputs from recent commercial generators. Experiments across multiple benchmarks show that combining frozen multimodal representations with carefully curated training data provides a simple and effective approach to AI-generated image detection.