💻 computer science

GlobalForge: Towards Robust AI-Generated Image Detection

The paper introduces GlobalForge, a robust AI-generated image detection framework that shifts focus from fragile local artifacts to stable global structural reasoning via a Local Information Bottleneck and Global Structural Reasoning module, significantly outperforming state-of-the-art methods on real-world degraded images and a newly proposed RealDeg-Bench.

Original authors: Manni Cui, Ruiqi Liu, Dianyuan Zou, Ziheng Qin, Jingrui Xu, ZiAn Wang, Jianglan Wei, Han Zhou, Yu Liu, Yan Wang, Shu Wu

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Manni Cui, Ruiqi Liu, Dianyuan Zou, Ziheng Qin, Jingrui Xu, ZiAn Wang, Jianglan Wei, Han Zhou, Yu Liu, Yan Wang, Shu Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to spot a forgery. In the world of digital images, "AI-generated" pictures are becoming so good that they look just like real photos to the naked eye. This is a big deal because bad actors could use these fake images to spread lies, create deepfakes of real people, or steal art. To fight back, scientists have built "detectors"—computer programs designed to sniff out the tiny, invisible mistakes that AI makes when it paints a picture.

For a long time, these detectors worked great in the lab. But there was a catch: they were like detectives who only knew how to spot a specific type of smudge on a pristine, untouched canvas. As soon as someone took that canvas, crumpled it, photocopied it, or shrank it down to fit on a phone screen (things that happen every time an image is shared online), the detectors got confused and failed. They were looking for "local artifacts"—tiny, fragile glitches in small patches of the image. When the image gets compressed or blurred, those tiny glitches vanish, and the detective is left blind. The big question was: how do we build a detector that doesn't panic when the evidence gets messy?

This is where the paper GlobalForge steps in. The researchers realized that the old detectives were focusing on the wrong clues. Instead of hunting for those tiny, easily-destroyed smudges, they decided to teach the AI to look at the "big picture"—the overall structure and how different parts of the image relate to each other from far away. They built a new framework called GlobalForge that forces the AI to ignore the fragile, local details and instead rely on the sturdy, global shape of the image.

Here is how they did it, using a few creative tricks:

First, they installed a "Local Information Bottleneck" (LIB). Think of this as a pair of special glasses that blur out the tiny, high-frequency details of the image. It's like telling the detective, "Don't look at the individual brushstrokes; they might be fake or erased. Just look at the general flow of the painting." By blurring these local details, the AI can't cheat by memorizing tiny, specific glitches that disappear when the image is compressed.

Second, they added a "Global Structural Reasoning" (GSR) module. Imagine you are trying to solve a puzzle, but you are forbidden from looking at pieces that are right next to each other. You have to jump across the table to see how a piece in the top-left corner connects to a piece in the bottom-right. This forces the AI to understand the long-range relationships in the image—the "global structure." If the image is AI-generated, these long-distance connections often feel "off" or unnatural, even if the local details are perfect.

To make sure this new detective was tough enough for the real world, the team created a new testing ground called RealDeg-Bench. Instead of just testing the AI on clean images or single types of damage, they simulated a "compound degradation chain." This is like taking a photo, compressing it, resizing it, blurring it, and then doing it all over again five times in a row—mimicking exactly what happens when a picture travels through social media.

The results were impressive. On these tough, real-world tests, GlobalForge maintained its accuracy, while older methods fell apart. Specifically, the new method improved the average balanced accuracy (a measure of how well it spots both real and fake) by 5.89% over the previous best methods on eight different "in-the-wild" benchmark groups. On their new RealDeg-Bench, which included complex chains of damage, GlobalForge reached an average balanced accuracy of 85.81%, clearly beating other top detectors.

The authors found that the old way of relying on local artifacts was indeed the root cause of the failures. When they blocked the AI from seeing those local clues, it actually got better at spotting fakes, even on images it had never seen before. They also showed that simply adding more "data augmentation" (showing the AI more damaged images during training) wasn't enough; the AI had to be forced to change how it looked at the image, not just what it saw.

In short, GlobalForge suggests that to catch AI fakes in the messy real world, we need to stop looking for tiny, fragile clues and start looking for the big, structural truth. It's a shift from being a detective who spots a smudge to one who understands the whole story.

Technical Summary: GlobalForge: Towards Robust AI-Generated Image Detection

Problem Statement

While AI-generated image (AIGI) detectors achieve high accuracy on clean benchmarks, their performance degrades sharply when images traverse real-world propagation channels (e.g., social media sharing). The paper identifies that this fragility stems from a fundamental flaw in current detection paradigms: detectors overfit to local artifacts (small spatial neighborhoods, high-frequency spectral traces, or upsampling grids) left by generative models. These local cues are easily destroyed by common degradations such as JPEG compression, blur, and resizing. Existing defenses, such as data augmentation or image restoration, are largely passive; augmentation fails to shift feature preferences toward unseen degradations, while restoration often pulls images back toward natural-image statistics rather than preserving generator-specific traces. Furthermore, current evaluation benchmarks often test single perturbations or rely on uncontrolled "in-the-wild" data, failing to capture the compound degradation chains typical of real-world propagation.

Methodology: GlobalForge

The authors propose GlobalForge, a detection framework designed to shift the discriminative cue from fragile local artifacts to robust global structural coherence. The framework consists of two complementary modules trained jointly under a degradation-aware contrastive structural loss:

Local Information Bottleneck (LIB):
- Goal: Suppress local high-frequency components to prevent the model from learning "shortcut" local artifacts.
- Mechanism: Instead of applying low-pass filtering to the raw input (which can introduce side effects like ringing), LIB operates in the feature domain. It takes deep features from the backbone and fuses the original features with a smoothed version (obtained via depthwise Gaussian convolution) using a learnable scalar gate ( $\alpha$ ). This controlled fusion attenuates high-frequency noise and local grid traces while preserving lower-frequency structural information.
Global Structural Reasoning (GSR):
- Goal: Force the model to rely on long-range relationships rather than local correlations.
- Mechanism: Built upon the output of LIB, GSR modifies the self-attention mechanism. It applies a local masking window (Chebyshev distance) to the attention score matrix, setting attention weights between a query token and its immediate neighbors to $-\infty$ . This forces every token to gather evidence exclusively from distant regions, compelling the model to integrate cross-region structural information for its decision.
Synergistic Optimization (LDCS):
- The modules are trained using a Degradation-Aware Contrastive Structural Loss (LDCS). For each input, an online compound-degraded view is generated. The loss function (based on InfoNCE) aligns the global image representations of the clean and degraded views in the feature space. This encourages the model to preserve degradation-invariant structural information, ensuring stability under realistic propagation.

Key Contributions

Diagnosis of Fragility: The paper provides a controlled diagnosis identifying local-artifact dependency as the root cause of degradation fragility. Experiments demonstrate that blocking access to local artifacts improves both degradation robustness and cross-generator generalization.
GlobalForge Framework: To the authors' knowledge, this is the first AIGI detection framework that actively suppresses local shortcuts and forces discriminative decisions onto long-range structural cues, decoupling robustness from generator-specific local signatures.
RealDeg-Bench: The authors introduce a new evaluation benchmark covering 7 common degradation operators and multi-step compound chains (up to 5 steps). This benchmark addresses the limitations of prior work by simulating realistic, controllable propagation depths and strengths.
Performance Gains: GlobalForge achieves state-of-the-art results, improving average Balanced Accuracy (BAcc) by 5.89% over the previous SOTA on 8 in-the-wild benchmark groups and reaching 85.81% average BAcc on RealDeg-Bench.

Experimental Results

The evaluation was conducted on two suites: RealDeg-Bench (controllable degradations) and In-the-Wild Benchmarks (8 groups including Chameleon, WildRF, and real-chain).

RealDeg-Bench: GlobalForge (variants Ours-d2 and Ours-d3) achieved the highest average BAcc (85.81%) across Clean, Multi-Kind, and Compound Degradation conditions. Unlike baselines (e.g., DDA, GAPL, B-Free) which exhibited "cliff-like" performance drops as degradation strength or chain length increased, GlobalForge maintained a smooth decay curve, confirming the stability of global structural features.
In-the-Wild Evaluation: On 17 test subsets across 8 benchmark groups, GlobalForge achieved an average BAcc of 85.93%, outperforming the strongest baseline (DDA) by 5.89%. It ranked first on 3 of the 8 groups and maintained consistent performance across diverse platforms and generators, whereas several baselines collapsed on out-of-distribution subsets.
Ablation Studies: Removing any core component (LIB, GSR, or LDCS) resulted in significant performance drops. The combination of LIB and GSR proved particularly effective, with LIB suppressing local shortcuts and GSR enforcing long-range reasoning. Alternative analyses showed that input-side filtering (pixel or frequency domain) was inferior to the feature-domain bottleneck, and unconstrained global attention was less effective than the masked local-window approach.

Significance and Claims

The paper claims that the fragility of current AIGI detectors is a structural issue rooted in their reliance on local artifacts. By intervening in the learning process to suppress these artifacts and enforce global structural reasoning, GlobalForge achieves a robustness profile that is decoupled from specific generator signatures. The authors position GlobalForge not merely as a performance improvement, but as a paradigm shift from defending fragile local cues to leveraging stable global structures. The introduction of RealDeg-Bench is highlighted as a critical step toward fine-grained diagnosis of detector robustness under realistic, compound degradation chains, moving beyond the limitations of single-perturbation or uncontrolled "in-the-wild" evaluations.