Frequency-Aware Semantic Fusion with Gated Injection for AI-generated Image Detection

1. Problem Statement

The rapid advancement of AI-generated images (AIGI) has created highly realistic visual content that challenges the reliability of digital media. While Vision Foundation Models (VFMs) (e.g., DINOv3) offer strong semantic representations and frequency-based methods capture subtle generative artifacts, existing approaches that combine these modalities suffer from limited generalization. Specifically, they perform well on "seen" generators but degrade significantly on "unseen" ones.

The authors identify two primary causes for this failure:

1. Frequency Shortcut Bias: Models tend to over-rely on easily distinguishable, generator-specific frequency cues during training. These cues do not generalize across different generative models, leading to poor performance under distribution shifts.
2. Cross-Domain Representation Conflict: Existing dual-stream frameworks often use direct fusion (e.g., concatenation or cross-attention) to combine low-level frequency features with high-level semantic features. Because these features reside in inherently different representation spaces, direct fusion creates mixed representations that degrade detection performance.

2. Methodology: FGINet

The authors propose the Frequency-aware Gated Injection Network (FGINet), a dual-stream framework designed to learn generalizable frequency representations and integrate them hierarchically with semantic features. The architecture consists of three core components:

A. Band-Masked Frequency Encoder (BMFE)

To mitigate frequency shortcut bias, BMFE learns generator-invariant frequency representations.

• Haar Wavelet Decomposition: The input image is decomposed into four sub-bands ($LL, LH, HL, HH$). The low-frequency $LL$ band is discarded as it contains redundant semantic information, while the high-frequency bands ($LH, HL, HH$) are concatenated.
• Band-Level Independent Masking: During training, the encoder applies random masking to non-overlapping patches within each high-frequency band. This perturbation forces the model to rely on diverse frequency patterns rather than specific generator artifacts, encouraging robustness.
• Token Encoding: The masked high-frequency map is processed through a lightweight convolutional stem, pooled, and projected into a compact frequency token ($T_{freq}$) compatible with the Transformer embedding space.

B. Layer-wise Gated Frequency Injection (LGFI)

To resolve the representation conflict, LGFI replaces direct fusion with a progressive, hierarchical injection mechanism.

• Semantic Backbone: The semantic pathway uses a pre-trained DINOv3 backbone adapted via LoRA (Low-Rank Adaptation) to minimize parameter overhead while preserving generalizable semantic priors.
• Gated Injection: Instead of fusing features at the end, the frequency token $T_{freq}$ is progressively injected into the class token of the Transformer blocks at every layer ($l$).
• Adaptive Gating: A learnable gate parameter $\alpha^{(l)}$ dynamically controls the contribution of the frequency token at each layer:
  $$\hat{t}^{(l)}_{cls} = t^{(l)}_{cls} + \alpha^{(l)} T_{freq}$$
  This allows the model to adaptively weigh frequency cues according to the hierarchical abstraction level, preserving the pre-trained semantic structure while integrating artifact cues.

C. Hyperspherical Compactness Learning (HCL)

To enhance generalization to unseen generators, the final classification stage employs a Hyperspherical Compactness Learning framework based on the CosFace objective.

• Hyperspherical Normalization: Features and classifier weights are normalized to lie on a unit hypersphere ($\|x\|=1$).
• Cosine Margin: An additive cosine margin is enforced during training to maximize inter-class separation and minimize intra-class variance. This creates compact, well-separated clusters for real and fake samples, improving decision boundaries.

3. Key Contributions

1. FGINet Framework: A novel architecture that addresses frequency shortcut bias and cross-domain conflicts via Band-Masked Frequency Encoding and Layer-wise Gated Injection.
2. BMFE Mechanism: Introduces cross-band independent masking to disrupt reliance on generator-specific cues, promoting diverse and generalizable frequency representations.
3. LGFI Mechanism: Proposes a progressive, gated injection strategy that aligns frequency cues with the hierarchical abstraction of VFMs, avoiding the pitfalls of direct fusion.
4. HCL Objective: Integrates hyperspherical constraints with cosine margins to enforce compact and separable feature distributions, boosting robustness against unseen generators.

4. Experimental Results

The authors evaluated FGINet on multiple challenging datasets, including GenImage, SynthBuster, WildRF, Chameleon, and RRDataset (for robustness).

• State-of-the-Art Performance:
  • On GenImage (8 generators), FGINet achieved 96.7% mean accuracy, outperforming the second-best method (SAFE) by 1.2%.
  • On SynthBuster (9 diffusion models), it achieved 94.3% mean accuracy, surpassing the runner-up (DDA) by 4.2%.
  • On WildRF (social media data), it achieved 95.9% accuracy, demonstrating strong cross-source generalization.
  • On the difficult Chameleon dataset (hard samples), FGINet reached 92.5% overall accuracy, significantly outperforming DDA (84.8%) and other SOTA methods.
• Robustness: In the RRDataset evaluation (simulating transmission and re-digitization), FGINet maintained superior performance compared to competitors, showing resilience against severe real-world degradations.
• Ablation Studies:
  • Removing BMFE reduced performance, confirming the need to disrupt frequency shortcuts.
  • Replacing LGFI with direct fusion (Concat/Add) or ungated injection caused significant performance drops, validating the necessity of hierarchical, gated integration.
  • Removing HCL reduced accuracy from 92.5% to 90.6%, highlighting the importance of hyperspherical constraints.

5. Significance

This paper makes a significant contribution to the field of AI-generated image detection by shifting the paradigm from simple feature fusion to structured, hierarchical integration.

• Generalization: It directly addresses the "unseen generator" problem, a critical bottleneck in real-world deployment, by preventing models from memorizing specific generator artifacts.
• Architectural Insight: The work demonstrates that frequency cues should not be treated as a separate stream to be fused late, but rather as a complementary signal that should progressively guide semantic representation learning through adaptive gating.
• Practical Impact: The strong performance on "in-the-wild" datasets (social media, re-digitized images) suggests that FGINet is a viable solution for deploying robust detection systems in real-world scenarios where image quality varies and generators are unknown.