Degradation-Consistent Paired Training for Robust AI-Generated Image Detection
Authors: Zongyou Yang, Yinghan Hou, Xiaokun Yang
Organizations: Department of Earth Science and Engineering, Imperial College London, London, United Kingdom · School of Electronic Information, Nanchang Institute of Technology, Nanchang, China
AI-generated image detectors suffer significant performance degradation under real-world image corruptions such as JPEG compression, Gaussian blur, and resolution downsampling. We observe that state-of-the-art methods, including B-Free, treat degradation robustness as a byproduct of data augmentation rather than an explicit training objective. In this work, we propose Degradation-Consistent Paired Training (DCPT), a simple yet effective training strategy that explicitly enforces robustness through paired consistency constraints. For each training image, we construct a clean view and a degraded view, then impose two constraints: a feature consistency loss that minimizes the cosine distance between clean and degraded representations, and a prediction consistency loss based on symmetric KL divergence that aligns output distributions across views. DCPT adds zero additional parameters and zero inference overhead. Experiments on the Synthbuster benchmark (9 generators, 8 degradation conditions) demonstrate that DCPT improves the degraded-condition average accuracy by 9.1 percentage points compared to an identical baseline without paired training, while sacrificing only 0.9% clean accuracy. The improvement is most pronounced under JPEG compression (+15.7% to +17.9%). Ablation further reveals that adding architectural components leads to overfitting on limited training data, confirming that training objective improvement is more effective than architectural augmentation for degradation robustness.
Figures & tables
Fig. 1: Comparison of (a) our baseline training and (b) DCPT training under the frozen-backbone setting. Both use the same frozen DINOv2 ViT-B/14 backbone and trainable classifier head (197K params). DCPT adds two consistency constraints ( Lfeat and Lpred ) between clean and degraded view representations, with zero additional parameters. Note: the original B-Free [ 1 ] fine-tunes the backbone end-to-end; our frozen setting is adopted to isolate the training objective contribution.
Method
Prm
Cln
J70
J50
J30
B1
B2
B3
R.5
R.25
Avg
B-Free ∗ [ 1 ]
ACC
86.5M
90.1
84.8
82.8
80.1
85.4
84.7
83.0
82.4
77.0
82.5
AUC
98.7
97.4
96.3
94.8
98.5
97.7
95.5
98.0
92.9
96.4
Baseline
ACC
197K
86.2
56.8
53.3
52.3
85.3
77.2
52.3
77.0
57.7
64.0
AUC
89.1
79.8
77.7
75.3
88.0
84.7
70.1
84.2
68.1
78.5
DCPT
ACC
197K
85.3
72.5
71.2
69.5
84.4
78.8
59.9
80.8
68.0
73.1
AUC
87.9
80.5
79.4
77.6
87.1
84.0
71.2
83.8
71.4
79.4
TABLE I: Degradation robustness on Synthbuster (avg across 9 generators). Best within same-architecture in bold .
Fig. 2: JPEG compression robustness. DCPT maintains a significantly flatter accuracy curve than the baseline, with improvements widening as compression increases.
Fig. 3: Per-generator accuracy under JPEG QF=70. DCPT consistently improves detection across all 9 generators.
Variant
λf
λp
Clean
J70
J50
J30
JPEG Avg
Baseline
0
0
86.2
56.8
53.3
52.3
54.1
Lfeat only
0.5
0
84.6
71.7
71.4
69.0
70.7
Lpred only
0
0.1
84.8
74.7
72.9
71.0
72.9
DCPT (both)
0.5
0.1
85.3
72.5
71.2
69.5
71.1
+ SRM (1.6M)
0.5
0.1
78.1
50.7
56.4
65.8
57.6
TABLE II: Ablation of DCPT loss components (Synthbuster, JPEG conditions). All variants use frozen DINOv2 + classifier head (197K params unless noted).
Fig. 4: Clean vs. degraded accuracy comparison. DCPT nearly halves the clean-to-degraded performance gap.
Fig. 5: Per-degradation accuracy improvement of DCPT over baseline. JPEG compression benefits most.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
AI-generated image (AIGI) detectors achieve strong accuracy on clean benchmarks, but their performance drops sharply after images are propagated through real-world channels. We trace this fragility to what these detectors actually learn: they overfit to local artifacts left by generators in small spatial neighborhoods, which are easily destroyed by common propagation degradations such as JPEG compression and blur. Instead, we shift the discriminative cue from fragile local artifacts to more robust global structure. Building on this, we propose GlobalForge, a framework with two complementary modules. The Local Information Bottleneck (LIB) suppresses local components to block shortcut learning, while the Global Structural Reasoning (GSR) module forces every token to gather evidence from distant regions. Both modules are trained jointly under a contrastive structural loss based on degradation that keeps the resulting features stable under degradation. To support fine-grained robustness evaluation, we further introduce RealDeg-Bench, covering 7 common degradation operators and multi-step compound chains. GlobalForge improves average BAcc on 8 in-the-wild benchmark groups by 5.89% over the previous state-of-the-art, and is clearly ahead of representative baselines on RealDeg-Bench under both single and compound degradations. Code is available at https://anonymous.4open.science/r/GlobalForge-BE0F/.
Manni Cui, Ruiqi Liu, Dianyuan Zou +8
1Huazhong University of Science and Technology · Institute of Automation, Chinese Academy of Sciences · 3Jilin University +1
While existing AI-generated image detectors report high performance, we identify that this is largely driven by a critical prediction asymmetry: a bias toward the real class that severely limits sensitivity to generated content, especially under standard post-processing operations such as compression and resizing. We hypothesize that this stems from the model's reliance on spurious features, distracting signals that obscure true generative artifacts. To address this, we propose DEAR (Dissect and Prune), which leverages inpainted images to identify and prune these interfering components. Specifically, we find that features strongly aligned to either inpainted or non-inpainted regions are less robust to post-processing. By measuring the alignment between channel activations and inpaint masks, DEAR removes features at both extremes, retaining only those that capture genuine generative artifacts. Experimental results demonstrate that our approach significantly enhances robustness against unseen generators and post-processing, effectively mitigating the prediction asymmetry. Our code is available at https://github.com/dahyedahye/dear.
Dahye Kim, Jaehyun Choi, Hyun Seok Seong +4
Korea AI Safety Institute (AISI), ETRI, Seongnam, South Korea · Department of Artificial Intelligence, Sungkyunkwan University, Suwon, South Korea
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