Beyond the Last Layer: Global Inter-Layer Aggregation for AI-Generated Image Detection

Authors: Wenbin Shen, Zhichao Lian

arXiv: 2611.00017 [cs.CV]

Submitted: 30 September 2026

Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)

DOI: https://doi.org/10.48550/arXiv.2611.00017

Abstract

The rapid development of generative models has made AI-generated images increasingly difficult to distinguish from real images. This progress poses several challenges for detectors, including rapid generator updates, substantial shifts in data distribution, and limited cross-dataset generalization. Many existing methods based on pretrained vision encoders rely primarily on the global representation from the last layer, potentially leaving useful information from intermediate layers underexplored.

To address this issue, we propose Global Inter-Layer Aggregation (GILA). GILA extracts global representations from different depths of a frozen encoder, models interactions among layer representations through inter-layer self-attention, and adaptively aggregates the resulting contextualized representations. The method is not tied to a single pretraining objective and does not require fine-tuning the pretrained encoder.

We evaluated GILA on three vision backbones with different pretraining objectives, ten seen generators, and three external datasets. Experimental results show that GILA makes better use of multi-layer information in pretrained encoders and improves both detection performance and cross-dataset generalization.

Key Contributions

Inter-layer representation use: Rather than relying solely on the final-layer global representation, GILA draws global features from multiple depths of a frozen pretrained encoder.
Inter-layer self-attention: Layer-wise representations are contextualized by modeling their interactions, enabling adaptive aggregation of complementary information.
Backbone-agnostic design: The approach is independent of any single pretraining objective and requires no fine-tuning of the pretrained encoder.
Strong empirical results: Evaluation across three vision backbones, ten seen generators, and three external datasets demonstrates improved detection accuracy and cross-dataset generalization.

Methodology Overview

GILA operates in three principal stages:

Multi-depth feature extraction — Global representations are collected from different depths of a frozen vision encoder.
Inter-layer self-attention — Interactions among the layer representations are modeled to produce contextualized features.
Adaptive aggregation — The contextualized representations are combined adaptively to form the final detection representation.

This design preserves the efficiency of frozen encoders while exploiting intermediate-layer information that conventional last-layer approaches discard.

Experimental Evaluation

The method was assessed along three dimensions:

Backbones: Three vision backbones with differing pretraining objectives.
Generators: Ten seen generators used during evaluation.
Generalization: Three external datasets to test cross-dataset performance.

Results indicate that GILA leverages multi-layer information more effectively than last-layer baselines, yielding gains in detection performance and generalization to unseen data distributions. As of 2026, with generative models advancing rapidly and producing increasingly photorealistic outputs, methods that generalize across generators and datasets—such as GILA—are of growing practical importance for content authenticity and provenance verification.

Citation

@article{shen2026gila,
  title   = {Beyond the Last Layer: Global Inter-Layer Aggregation for AI-Generated Image Detection},
  author  = {Shen, Wenbin and Lian, Zhichao},
  journal = {arXiv preprint arXiv:2611.00017},
  year    = {2026},
  doi     = {10.48550/arXiv.2611.00017}
}

Access

Abstract page: https://arxiv.org/abs/2611.00017
PDF (v1): https://arxiv.org/abs/2611.00017v1
DOI: https://doi.org/10.48550/arXiv.2611.00017

Submission History

[v1] Wed, 30 Sep 2026 13:14:38 UTC (1,388 KB)

via ArXiv CV

Beyond the Last Layer: Global Inter-Layer Aggregation for AI