The proliferation of highly realistic AI-Generated Image (AIGI) has necessitated the development of practical detection methods. While current AIGI detectors perform admirably on clean datasets, their detection performance frequently decreases when deployed "in the wild", where images are subjected to unpredictable, complex distortions. To resolve the critical vulnerability, we propose a novel LoRA-based Pairwise Training (LPT) strategy designed specifically to achieve robust detection for AIGI under severe distortions. The core of our strategy involves the targeted finetuning of a visual foundation model, the deliberate simulation of data distribution during the training phase, and a unique pairwise training process. Specifically, we introduce distortion and size simulations to better fit the distribution from the validation and test sets. Based on the strong visual representation capability of the visual foundation model, we finetune the model to achieve AIGI detection. The pairwise training is utilized to improve the detection via decoupling the generalization and robustness optimization. Experiments show that our approach secured the 3th placement in the NTIRE Robust AI-Generated Image Detection in the Wild challenge
Figures & tables
Figure 1: Previous AIGI detection methods mainly consider the generalization performance under different generation models. However, when the images are distorted due to the transmission, the detectors will make false judgments. To make the detector practical in the downstream scenario, it is urgent to consider a robust and generalized detector.
Figure 2: Distorted Images Illustration. Irrespective of whether the images are real or fake, all images are subjected to multiple distortions with high-level strength, such as Gaussian blur, quantization, impulse noise, brighteness, and tone curve transformations.
Dataset
Images
Generators
Distortions
Toy
10,000
10
0
Train
277,000
20
0
Validation-1
10,000
25
5
Validation-2
2,500
25
5
Public test
2,500
30
7
Private test
2,500
35
9
Table 1: Dataset statistics. The validation stage only involves simple distortions. The test stage further includes complex distortions. The distortion in Validation-2 is harder than in Validation-1.
Figure 3: (a) Image Area Statistics. There is a distribution mismatching between the training and validation/test sets. (b) Special Cases from Validation/Test sets. Some images only appear to have local semantic or non-semantic information.
Figure 4: The overall framework of LPT. To achieve robust detection, our strategy involves the simulation of data distribution, the fine-tuning of the visual foundation model, and the pairwise training, respectively.
Type
Distortions
Blur
Gaussian / Lens
Noise
Gaussian / Impulse / Speckle
Color
Shift/ ColorJitter / Moire/ Tone
Brightness
Brighten / Darken
Compression
JPEG / Quantization
Spatial Space
Jitter
Table 2: The introduction of distortions during the model training.
Participants
Public
Private
Test Hard Set
Test Hard Set
MICV
97.38
97.22
Ant International
97.30
97.20
INTSIG
90.78
91.30
vincentlc
86.42
87.33
Reagvis Labs
85.66
86.00
Table 3: Performance comparison on the challenge test set (AUC %). Our approach secures the third placement. ‘ Hard ’ denotes that the results are computed from the distorted images only.
LCE
LKL
LMSE
Validation*
Public Test
✓
-
-
97.31
87.91
✓
✓
-
97.38
88.06
✓
✓
✓
97.79
89.26
Table 4: Ablation study on training strategy (AUC %). Each ✓ indicates that the component is included. ‘*’ means the results are the average from Validation-1 and Validation-2 sets. All results are computed from the clean and distorted images.
Size
Distortion
Val.-1
Val.-2
Public Test
Avg.
-
-
98.89
96.69
89.26
94.95
✓
-
99.17
96.77
89.27
95.07
✓
✓
99.10
96.74
89.58
95.15
Table 5: Ablation study on the data distribution simulation (AUC %). All results are computed from the clean and distorted images.
Figure 5: Left : The impact of the distortion strength. Right : The impact of additional data incorporation.
Model
Validation-1
Validation-2
Public Test
CLIP*
99.50
98.05
92.67
Meta-CLIP2*
99.60
98.44
95.73
EVA-CLIP*
99.22
98.06
95.15
Table 6: Ablation study on the visual foundation model selection (AUC %). Meta-CLIP2 [ 3 ] is the Big-G version, and EVA-CLIP [ 32 ] is the Big-E version. All results are computed from the clean and distorted images. ‘*’ denotes the utilization of data extension.
Methods
Clean
JPEG
Noise
Blur
Avg.
Q=50
k=7 σ =2
v=25
NPR [ 34 ]
91.80
50.20
50.00
65.70
64.42
MLEP [ 41 ]
89.85
48.90
49.90
74.90
65.88
CNN [ 37 ]
74.91
57.57
70.48
74.80
69.44
Effort [ 39 ]
97.22
60.32
68.56
90.88
79.24
DDA [ 1 ]
84.16
78.02
79.65
83.40
81.30
Table 7: Performance comparison (Acc %) on the ForenSynths [ 37 ] . Except the ‘ Avg. ’ column, each result is the average detection accuracy from the eight generation methods (‘ProGAN’, ‘CycleGAN’, ‘Deepfake’, ‘GauGAN’, ‘BigGAN’, ‘StarGAN’, ‘StyleGAN’, and ‘StyleGAN2’). Here, Q, k, σ , and v denote quality factor, kernel size, stand derivation, and variation.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Australian National University, Canberra, Australia · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +1