PaperlayerView exact revision on arXiv

Source-linked AI summary

EIB-Net: Entropy-Guided Information Bottleneck for Generalizable AI-Generated Image Detection

Zhida Zhang, Xinlei Ma, Jie Cao

arXiv:2609.29064v1cs.CVcs.AI

TL;DR

AI-generated images lack the explicit traces that many existing forgery detectors target, and their local texture can be less faithful than their global semantics. EIB-Net selects a low-entropy patch and compresses its features; experiments report state-of-the-art accuracy across intra- and cross-generator settings, including over 85% accuracy with 2% of training data.

  • Problem

    Generation-based images lack the explicit traces that manipulation-focused detectors commonly use, making their detection challenging.

    Source: §I · p. 1

  • Method

    EIB-Net selects the lowest-entropy image patch with Image Entropy and applies VIB to compress features and discard content semantics.

    Source: §I · p. 1

  • Results

    EIB-Net reports state-of-the-art accuracy in intra- and cross-generator settings and over 85% accuracy with only 2% of training data.

    Source: §I · p. 1

  • Takeaways & Limitations

    EGPL is reported to consistently enhance diverse CNN and Transformer backbones.

    Source: §I · p. 2

Abstract

from arXiv · show

The proliferation of photorealistic AI-generated images demands robust detection methods that generalize across diverse generative models. While existing approaches target manipulation-based forgeries with local artifacts, generation-based images (e.g., from diffusion models) lack such traces, posing a fundamental challenge. We observe that generative models prioritize global semantics at the expense of local texture fidelity, making low-texture regions key indicators of synthetic origin. To exploit this, we propose EIB-Net, an Entropy-guided Information Bottleneck Network. EIB-Net introduces a novel Image Entropy (IE) metric to automatically select the most informative (lowest-entropy) patch, then processes it with a Variational Information Bottleneck (VIB) to learn compact, generalizable features. Extensive experiments on DIFF, DiffusionForensics, and GenImage benchmarks demonstrate state-of-the-art performance: EIB-Net achieves 85.7\% accuracy using only 2\% of training data, outperforming full-image baselines by over 15\%, and maintains robust cross-generator generalization (83.5\% average accuracy on GenImage). Furthermore, our entropy-guided patch selection (EGPL) consistently enhances diverse backbones (CNNs and Transformers), proving its practical value for data-efficient detection.

UCAS & NLPR, CASIA

Generation-based images lack the localized traces used by many forgery detectors, while generative models can sacrifice local texture fidelity for global semantic coherence. EIB-Net selects a smooth, artifact-prone patch and compresses its features to focus detection on local statistical cues.

I. INTRODUCTION

  • Manipulation-based detectors often rely on localized artifacts, but generated images lack explicit traces and their flaws can be distributed across the image.Generative models prioritize global semantic coherence at the expense of local texture fidelity, particularly in low-texture regions.Source: §I · p. 1
  • EIB-Net uses a “select-then-compress” strategy: Image Entropy selects the lowest-entropy patch, then VIB discards content semantics to retain compact authenticity-relevant features.The framework focuses on artifact-prone regions rather than dataset-specific semantics.Source: §I · pp. 1–2
  • EIB-Net reports state-of-the-art accuracy across intra- and cross-generator settings, with over 85% accuracy using only 2% of training data.The EGPL strategy is also reported to enhance diverse backbone architectures.Source: §I · pp. 1–2
  • The Image Entropy metric selects smooth, artifact-prone patches, with the stated aim of improving convergence speed and generalization in low-data regimes.The paper presents EGPL as applicable to CNN and Transformer backbones.Source: §I · p. 2

II. RELATED WORK

Prior generated-image detectors include reconstruction, autoencoder, and statistical approaches, while the paper distinguishes its focus on low-entropy regions. EIB-Net pairs that patch selection with VIB compression to reduce reliance on semantic content.

A. Generated Image Detection

  • Prior manipulation-focused detectors rely on boundary or lighting artifacts –, while photorealistic latent-diffusion images lack explicit tampering traces.Other approaches use diffusion reconstruction, autoencoder representations,, or statistical and structural cues,.Source: A. Generated Image Detection · p. 2
  • Unlike approaches that rely on reconstruction, geometric inconsistencies, or global semantic priors, EIB-Net targets low-texture, low-entropy regions where synthesis artifacts are expected.The paper contrasts this choice with camera-based forensics, which typically selects texture-rich patches for device-specific noise.Source: A. Generated Image Detection · p. 2

B. Mutual Information and Information Bottleneck

  • The Information Bottleneck compresses input X into latent Z while preserving task-relevant information about target Y; EIB-Net’s VIB discards content information from the selected patch.The paper frames patch selection and feature compression as a dual strategy operating at data and feature levels.Source: B. Mutual Information and Information… · pp. 2–3

A. Overview of EIB-Net

  • EIB-Net selects the smoothest patch with EGPL, then processes it through a ResNet encoder and VIB layer to retain statistical cues for authenticity classification.The paper describes this “select-then-compress” design as reducing reliance on distracting global semantics.Source: A. Overview of EIB-Net · p. 3

1) Image Entropy Computation:

EIB-Net measures local texture complexity to select the smoothest patch, then compresses its features into a compact latent representation for authenticity detection.

1) Image Entropy Computation:

  • Image Entropy sums multidirectional intensity-gradient variation across color channels and patch locations; lower values indicate smoother, less-textured regions.The metric uses neighboring and diagonal directions to quantify local texture complexity.Source: 1) Image Entropy Computation · p. 3

2) Patch Selection:

  • EGPL selects the minimum-entropy patch as the detector’s sole input, following the hypothesis that low-entropy areas are weak points for generators.The selected region is intended to concentrate detection on areas where generative supervision is likely weakest and anomalies most pronounced.Source: 2) Patch Selection · p. 3

C. Network Architecture with VIB

  • The VIB layer maps the selected patch’s ResNet features into a latent code designed to retain authenticity information while discarding residual semantics.The code is sampled from a Gaussian distribution using the reparameterization trick.Source: C. Network Architecture with VIB · p. 3

D. Loss Function

The loss combines classification and prediction consistency with information-bottleneck regularization to balance accuracy, consistency, and feature compression.

D. Loss Function

  • The composite loss balances classification accuracy, prediction consistency, and feature compression.Figure 2 depicts the pipeline from minimum-entropy patch selection through ResNet encoding and VIB compression to real/synthetic classification.Source: D. Loss Function · p. 3, Figure 2 · p. 4
  • The classification loss combines label-smoothed cross-entropy with Jensen-Shannon divergence regularization, weighted by α = 12.The smoothing factor is ϵ = 0.1; the paper says it prevents over-confidence, while the divergence term encourages consistent output distributions.Source: D. Loss Function · p. 4
  • The VIB is regularized by the KL divergence between the learned latent distribution q(z|x) and a standard Gaussian prior p(z).The compression trade-off parameter β is set to 10^-3 in all experiments.Source: D. Loss Function · p. 4

A. Experimental Setup

Across the reported benchmarks, EIB-Net is presented as data-efficient and generalizable, including when tested on generators unseen during training.

A. Experimental Setup

  • The setup uses DIFF, DiffusionForensics, and GenImage, evaluating both same-generator and unseen-generator testing with accuracy as the reported metric.The DiffusionForensics setting trains on LSUN-Bedroom and tests on ImageNet samples from ADM and SDv1.Source: A. Experimental Setup · p. 4, 1) DiffusionForensics · p. 4

B. Intra-Model Generalization

  • EIB-Net is reported to outperform other methods across DIFF-Intra training-data regimes, especially in low-data conditions.The analysis attributes its stable, sample-efficient learning to EGPL’s focus on low-entropy, artifact-prone regions, contrasting it with full-image models’ overfitting to limited semantic cues.Source: Table I · p. 4, B. Intra-Model Generalization · p. 4

1) DiffusionForensics:

  • With 1% of the training data, EIB-Net maintains >99% validation accuracy and >80% test accuracy across ADM and SDv1 in the cross-scene, cross-model setup.The authors contrast this with baselines that suffer a catastrophic test-performance drop despite strong validation results; Table II presents the cross-model results by training ratio.Source: 1) DiffusionForensics · pp. 4–5, Table II · p. 5

2) GenImage:

  • 83.51% average accuracy across unseen generators and 99.89% on SD v1.4 are reported for EIB-Net on GenImage.Methods were trained exclusively on SD v1.4 images and tested on samples from eight different generators.Source: 2) GenImage · p. 5

D. Ablation Study

  • The ablation study examines the contributions of EGPL and VIB to EIB-Net’s design.The supplied passage states the study’s focus but does not report ablation outcomes.Source: D. Ablation Study · p. 5

1) Analysis of EGPL Component:

EGPL favors low-entropy patches, with 16-patch splitting offering a useful balance; the full EIB-Net combines patch selection with VIB, and EGPL benefits multiple backbone architectures. The conclusion reports strong cross-generator performance with little training data.

1) Analysis of EGPL Component:

  • On a 16-patch split, minimum-entropy selection matches full-image accuracy with faster convergence, while maximum-entropy selection performs worst.The figure reports validation accuracy for ResNet50, smoothed over 20 epochs.Source: 1) Analysis of EGPL Component · p. 5, Figure 4 · p. 5
  • Sixteen patches offer the best granularity trade-off; too few retain irrelevant content, too many fracture artifact signals, and one full-image patch loses information through rescaling.The patch-count comparison uses the lowest-entropy patch from each split.Source: 1) Analysis of EGPL Component · p. 5, Figure 4 · p. 5
  • EGPL improves performance across CNNs and Transformers on intra- and cross-model benchmarks, including ResNet50, ViT-B16, and RMT-S.Table IV compares EGPL’s impact across backbone architectures and datasets.Source: 1) Analysis of EGPL Component · p. 5, Table IV · p. 6

2) Analysis of VIB and Integrated Framework:

V. CONCLUSION

  • The conclusion describes EIB-Net as a select-then-compress framework and reports state-of-the-art cross-generator generalization with over 85% accuracy using 2% of training data.Source: §V · p. 6

I. DATASET DETAILS

The supplementary material describes three public benchmarks and a balanced DIFF-Intra subset, then specifies cross-model data splits and baseline backbone training configurations.

A. Overview of Evaluation Datasets

  • The evaluation uses DiffusionForensics, GenImage, and DIFF, covering diffusion-generated images, multiple generators, and face-focused data.The passage identifies dataset contents and printed bibliography markers: DiffusionForensics, GenImage, and DIFF.Source: A. Overview of Evaluation Datasets · p. 7

B. Construction of the DIFF-Intra Subset

C. Data Splits for Cross-Model Evaluation •

  • DiffusionForensics trains on LSUN-Bedroom ADM images and tests on ImageNet samples from ADM and SDv1; GenImage trains on SD v1.4 and evaluates eight generators.The DiffusionForensics split uses 80,000 real and 80,000 ADM-generated training images; GenImage uses 200k training images and 20k per generator for evaluation.Source: C. Data Splits for Cross-Model… · p. 7

A. Backbone Architectures and Baseline Training

  • EGPL is evaluated by integrating the module into four diverse, off-the-shelf backbone networks with minimal modification; all baselines replace only the final classification layer, avoiding extensive backbone-specific tuning.Source: A. Backbone Architectures and Baseline… · p. 7

B. Configuration for EIB-Net (EGPL + VIB)

The EIB-Net configuration specifies its ResNet50-based training setup, loss, and VIB size. Its implementation selects a low-entropy patch, extracts and compresses features, and computes Image Entropy from directional texture variation.

B. Configuration for EIB-Net (EGPL + VIB)

B. Implementation of Image Entropy

  • The pipeline divides each image into non-overlapping patches, selects and resizes the minimum-entropy patch, then extracts and compresses its features with the VIB layer.Source: B. Implementation of Image Entropy · pp. 7–8
  • Image Entropy sums absolute responses from directional Laplacian filters across spatial locations and RGB channels; lower values indicate smoother, less textured patches.The design is intended to capture fine-grained, channel-wise texture anomalies.Source: B. Implementation of Image Entropy · pp. 7–8

A. Comprehensive Analysis of Patch Selection

The patch-selection study finds that minimum-entropy selection with 16 patches performs best on DIFF-Intra, while random patches remain competitive and maximum-entropy patches are less effective.

A. Comprehensive Analysis of Patch Selection

B. Sensitivity Analysis of the VIB Weight β

The ablation selects β = 10−3 for the main experiments and finds accuracy gains from both label smoothing and JSD consistency regularization.

B. Sensitivity Analysis of the VIB Weight β

  • The β sensitivity study trained full EIB-Net on DIFF-Intra with EGPL selecting the minimum-entropy patch from a 16-patch split.It evaluated β values of 10−4, 10−3, and 10−2 while holding other hyperparameters constant.Source: B. Sensitivity Analysis of the VIB… · p. 9
  • 97.17% validation accuracy at β = 10−3 was the best result, versus 96.75% at β = 10−4 and 96.86% at β = 10−2.The authors attribute the lower scores to under-compression at β = 10−4 and likely excessive information loss at β = 10−2; β = 10−3 was used for the main experiments.Source: B. Sensitivity Analysis of the VIB… · p. 9, Table II · p. 9

C. Ablation on Classification Loss Components

  • +1.57% total gain over CE: label smoothing adds +0.65%, and JSD consistency regularization adds a further +0.88%.The full JSD loss yields the best performance in the progressive ablation.Source: Table III · p. 9, C. Ablation on Classification Loss… · p. 9
  • The authors describe JSD consistency regularization as integral to combating semantic overfitting and learning robust, generalizable features.They connect the regularizer’s self-consistency with the goal of retaining authenticity-relevant cues.Source: C. Ablation on Classification Loss… · p. 9

Related papers

Loading 2609.29064v1…