[Paper] Aggregating Diverse Cue Experts for AI-Generated Image Detection

Source: arXiv - 2601.08790v1

Overview

The paper presents Multi‑Cue Aggregation Network (MCAN), a detection framework that fuses several complementary signals—spatial content, high‑frequency edge details, and a novel chromatic‑inconsistency cue—to spot AI‑generated images. By treating these cues as “experts” and letting a mixture‑of‑encoders dynamically weight them, MCAN achieves markedly better cross‑model generalization than prior detectors that rely on a single type of feature.

Key Contributions

• Unified multi‑cue architecture that jointly processes spatial, frequency‑domain, and chromaticity information within a single network.
• Mixture‑of‑encoders adapter that learns to select and combine cue‑specific representations on the fly, improving robustness to unseen generators.
• Chromatic Inconsistency (CI) cue, which normalizes intensity and isolates acquisition‑noise patterns that differ between real photographs and synthetic outputs.
• State‑of‑the‑art performance on three major benchmarks (GenImage, Chameleon, UniversalFakeDetect), with up to 7.4 % absolute accuracy gain over the previous best method on GenImage.
• Extensive ablation studies demonstrating the individual and combined impact of each cue and the adaptive encoder mixture.

Methodology

1. Cue Extraction

   • Image cue: the raw RGB image, preserving overall scene semantics.
   • High‑frequency cue: obtained via a Laplacian filter (or wavelet transform) to highlight edges and fine textures that synthetic models often mishandle.
   • Chromatic Inconsistency cue: the image is first intensity‑normalized; residual chromatic variations (color‑channel noise) are then extracted, exposing subtle artifacts left by the generation pipeline.

2. Mixture‑of‑Encoders Adapter

   • Each cue is fed into its own lightweight encoder (e.g., ResNet‑18 blocks).
   • A gating network predicts a set of mixture weights conditioned on the input, effectively deciding how much each encoder’s output should contribute for a given image.
   • The weighted encoder outputs are concatenated and passed through a shared classifier head that outputs a real‑vs‑synthetic probability.

3. Training & Loss

   • Standard binary cross‑entropy loss with label smoothing.
   • An auxiliary contrastive loss encourages the network to keep cue‑specific embeddings discriminative across real and fake samples.

4. Implementation Details

   • Trained on a balanced mix of real photographs and AI‑generated images from eight popular generators (Stable Diffusion, DALL·E, Midjourney, etc.).
   • Data augmentation includes random cropping, JPEG compression, and color jitter to simulate real‑world distribution shifts.

Results & Findings

• Cross‑generator robustness: MCAN maintains >90 % accuracy even on generators it never saw during training, confirming the benefit of cue diversity.
• Ablation: Removing the CI cue drops accuracy by ~2.5 %; dropping the mixture‑of‑encoders (using simple concatenation) reduces performance by ~3 %, highlighting both components’ importance.
• Efficiency: The full model runs at ~45 ms per 512×512 image on a single RTX 3080, making it viable for real‑time moderation pipelines.

Practical Implications

• Content moderation platforms can integrate MCAN to flag synthetic media with higher confidence, reducing false positives that plague single‑cue detectors.
• Digital forensics tools gain a more reliable “expert system” that works across emerging generative models without needing frequent retraining.
• Social media APIs can expose a lightweight MCAN endpoint for developers to pre‑screen user uploads, helping combat misinformation and deep‑fake scams.
• Enterprise security: MCAN’s fast inference enables on‑device or edge deployment (e.g., in browsers or mobile apps) to detect AI‑generated images before they reach servers, preserving bandwidth and privacy.

Limitations & Future Work

• Cue selection bias: The current cues are handcrafted; future work could explore learnable cue discovery (e.g., via attention over spectral bands).
• Domain shift: While MCAN generalizes well across generators, extreme post‑processing (heavy stylization, aggressive compression) still degrades performance.
• Scalability to video: Extending the multi‑cue paradigm to temporal data (frame‑wise and motion cues) is an open direction.
• Explainability: The mixture weights provide some interpretability, but deeper analysis of why certain cues dominate for specific images would aid trustworthiness.

Bottom line: MCAN shows that aggregating diverse, complementary “expert” cues—spatial, frequency, and chromatic—offers a practical, high‑performing solution for AI‑generated image detection, ready for integration into today’s content‑safety stacks.

Authors

• Lei Tan
• Shuwei Li
• Mohan Kankanhalli
• Robby T. Tan

Paper Information

• arXiv ID: 2601.08790v1
• Categories: cs.CV
• Published: January 13, 2026
• PDF: Download PDF