cs.CVAug 12, 2026

Do CNNs Internally Represent Real and Fake Images Differently? A Hidden-Layer Analysis

Authors: Moumita Sen Sarma, Pascal Hitzler, Eugene Y. Vasserman

Organizations: Department of Computer Science, Kansas State University, Manhattan, Kansas, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Fake/synthetic images are increasingly prevalent, but it remains unclear whether Convolutional Neural Networks (CNNs) process real and fake images in the same internal manner. This work examines the hypothesis that CNNs represent real and fake images differently, such that fake images induce different hidden-layer activation patterns even when semantic content is preserved. The hypothesis is evaluated in scene recognition settings using trained CNN models. Dense-layer activations are extracted, and neurosymbolic methods assign semantic labels to selected neurons. For each real test image, corresponding fake images are generated with similar semantic content using object-label-guided text-to-image and image-to-image generation based on Stable Diffusion variants. Paired real-fake activation patterns are then compared statistically. Additional experiments with another dataset, CNN architecture, generative model, and JPEG/blur degradation analysis assess robustness. Results suggest that fake images evoke different hidden-neuron activations, and these differences are not explained only by simple image degradation. Overall, the findings indicate that real and fake images differ in CNN hidden-layer activation behavior at least in some settings, which opens the door for follow-up work on making use of this different behavior to improve fake image detection.

Figures & tables

Figure 1: Stable Diffusion Model Architecture. Adapted from Ritika (2023)
Figure 2: Workflow for comparing real and generated images through hidden-layer neuron activations. In this approach, fake images are generated using both text-only and image along with text settings before activation-level statistical comparison.
Figure 3: Sample fake output images using SUN2012 and Stable Diffusion variants; (a): from Text-Guided Fake Image Generation using ground-truth object annotations; (b): from Structure-Guided Fake Image Generation. Additional examples are depicted in Figure 11 in Appendix D.1 .
HypothesisNeuron SubsetSelection CriteriaMetric ComparedInterpretation
H1AllActivated ( ≥0.8×max⁡\geq 0.8\times\max ), zero-filteredCountReal images trigger broader neural activity than fake
H2Relevant confirmedi) Semantically aligned or ii) activatedValueReal images activate relevant neurons more strongly
H3Activated for realActivation thresholdShared CountFake images fail to preserve real activation patterns
H4Activated & confirmedConfirmed with activation thresholdCountReal images better activate semantically meaningful neurons
Table 1: Summary of the filtering criteria and interpretation for the four hypotheses.
Neuron IDECII ConceptsTLA %Non-TLA %Target MedianNon-Target MedianTarget MeanNon-Target Meanzz -scorepp -value
0snowy_mountain9554.447.050.256.121.04-6.57<0.00001<0.00001
7sky_and_snowy_mountain9538.812.810.002.720.56-5.92<0.00001<0.00001
9fence_and_central10062.304.030.763.971.46-5.54<0.00001<0.00001
………………………<0.00001<0.00001
41skyscraper8069.921.211.032.051.46-1.400.15536\mathbf{0.15536}
………………………<0.00001<0.00001
Table 2: Statistical evaluation results of concepts of confirmed neurons from concept induction using Mann-Whitney U. Bold rows represent neurons with pp -value ≥0.05\geq 0.05 , where the null hypothesis cannot be rejected . The complete table is represented in Figure 7 in Appendix A .
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired793793
Nonzero diffs used402348
Median(diff)1.0001.000
Mean(diff)0.5420.787
Prop(real >> fake)0.6000.681
pp -value2.749528×10−62.749528\times 10^{-6}2.939991×10−122.939991\times 10^{-12}
Table 3: Hypothesis 1 statistical results with SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Generation (Concept Alignment)Structure-Guided Fake Generation (Concept Alignment)Text-Guided Fake Generation (Activation-Based Relevancy)Structure-Guided Fake Generation (Activation-Based Relevancy)
Images paired890885244249
Nonzero diffs used890885244249
Median (real - fake)0.70471.04292.53632.2209
Mean (real - fake)0.74211.20232.65522.5786
Prop(real - fake)0.6570.7400.9590.924
pp -value4.681×10−284.681\times 10^{-28}5.839×10−665.839\times 10^{-66}4.507×10−414.507\times 10^{-41}2.117×10−402.117\times 10^{-40}
Table 4: Hypothesis 2 statistical results with SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired793793
Nonzero diffs used299264
Mean(real_count)0.8030.803
Mean(fake_count)0.0570.158
Median (real_count - fake_count)1.0001.000
Prop(real_count >> fake_count)1.0001.000
Table 5: Hypothesis 3 statistical results with SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired793793
Nonzero diffs used212189
Median (real-fake) counts1.0001.000
Prop(real >> fake)0.6460.709
pp -value6.272851×10−66.272851\times 10^{-6}1.724670×10−81.724670\times 10^{-8}
Effect size ( rrb)r_{rb})0.2920.418
Table 6: Hypothesis 4 statistical results with SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model.
Figure 4: Suppression in neurons and their mean activation drop for SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model. (a) Text-guided fake image generation (all neurons), (b) Structure-guided fake image generation (all neurons), (c) Text-guided generation for confirmed neurons, and (d) Structure-guided generation for confirmed neurons. Complete table is represented in Appendix B .
Figure 5: Structural Distortion Effects in Structure-Guided Fake Image Generation
Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Neuron IDECII ConceptsTLA %Non-TLA %Target MedianNon-Target MedianTarget MeanNon-Target Meanzz -scorepp -value
0snowy_mountain9554.447.050.256.121.04-6.57<0.00001<0.00001
7sky_and_snowy_mountain9538.812.810.002.720.56-5.92<0.00001<0.00001
9fence_and_central10062.304.030.763.971.46-5.54<0.00001<0.00001
11bathtub10051.984.790.054.690.75-7.23<0.00001<0.00001
12coffee_and_bouquet9555.161.370.201.600.78-3.820.000060.00006
16skyscraper_and_building9545.952.680.002.450.59-5.57<0.00001<0.00001
Appendix
Table 7: Statistical evaluation results of concepts of confirmed neurons from concept induction. Bold rows represent neurons with pp -value ≥0.05\geq 0.05 , where the null hypothesis cannot be rejected .
Figure 6: Suppression in neurons and their mean activation drop for SUN2012 dataset, InceptionV3 CNN model, and Stable Diffusion image generation model. (a) Text-guided fake image generation (all neurons), (b) Structure-guided fake image generation (all neurons), (c) Text-guided generation for confirmed neurons, and (d) Structure-guided generation for confirmed neurons.
StatisticReal Images with JPEG Compression
Images paired793
Nonzero diffs used304
Median(real - compressed_real)-1.000
Mean(real - compressed_real)-0.280
Prop(real >> compressed_real)0.418
pp -value0.9972669
Appendix
Table 8: Hypothesis 1 statistical results for real images with JPEG compression
StatisticConcept AlignmentActivation-Based Relevancy
Images paired966260
Nonzero diffs used966260
Median(real - compressed_real)-0.03830.3315
Mean(real - compressed_real)-0.06560.4086
Prop(real - compressed_real >> 0)0.4780.677
pp -value0.96769.09×10−139.09\times 10^{-13}
Appendix
Table 9: Hypothesis 2 statistical results for real images with JPEG compression
StatisticReal Images with JPEG Compression
Images paired793
Nonzero diffs used168
Mean(real_count)0.803
Mean(compressed_real_count)0.480
Median(real_count - compressed_real_count)1.000
Prop(real_count >> compressed_real_count)1.000
Appendix
Table 10: Hypothesis 3 statistical results for real images with JPEG compression
StatisticReal Images with JPEG Compression
Images paired793
Nonzero diffs used151
Median(real - compressed_real) counts-1.000
Prop(real >> compressed_real)0.457
pp -value0.8666343
Effect size ( rrbr_{rb} )-0.086
Appendix
Table 11: Hypothesis 4 statistical results for real images with JPEG compression
Figure 7: Suppression in neurons and their mean activation drop for JPEG compression. (a) Results for all neurons, (b) Results for confirmed neurons.
StatisticMild BlurMedium Blur
Images paired793793
Nonzero diffs used295302
Median(real - blurred_real)1.0001.000
Mean(real - blurred_real)-0.0440.331
Prop(real >> blurred_real)0.5150.606
pp -value0.60903084.809483×10−44.809483\times 10^{-4}
Appendix
Table 12: Hypothesis 1 statistical results for real images with mild and medium blurred real images
StatisticMild BlurMedium Blur
 Concept AlignmentActivation-Based RelevancyConcept AlignmentActivation-Based Relevancy
N total967260960260
Nonzero diffs used967260960260
Median(real - blurred_real)0.02260.48660.17310.7375
Mean(real - blurred_real)0.02610.50520.18610.7686
Prop(real >> blurred_real)0.5080.7540.5710.762
Appendix
Table 13: Hypothesis 2 statistical results for real images with mild and medium blurred real images
StatisticMild BlurMedium Blur
Images paired793793
Nonzero diffs used189229
Median(real_count - blurred_real_count)1.0001.000
Prop(real_count >> blurred_real_count)1.0001.000
pp -value1.085851×10−351.085851\times 10^{-35}2.100855×10−422.100855\times 10^{-42}
Effect size ( rrbr_{rb} )1.0001.000
Appendix
Table 14: Hypothesis 3 statistical results for real images with mild and medium blurred real images
StatisticMild BlurMedium Blur
Images paired793793
Nonzero diffs used155162
Median(real - blurred_real)1.0001.000
Prop(real >> blurred_real)0.5230.605
pp -value0.38834561.639224×10−31.639224\times 10^{-3}
Effect size ( rrbr_{rb} )0.0450.210
Appendix
Table 15: Hypothesis 4 statistical results for real images with mild and medium blurred real images
Figure 8: Suppression in neurons and their mean activation drop. (a) Mild Blurred Image (all neurons), (b) Medium Blurred Image (all neurons), (c) Mild Blurred Image (confirmed neurons), and (d) Medium Blurred Image (confirmed neurons).
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used399339
Median(real-fake)1.0001.000
Mean(real-fake)0.5740.209
Prop(real >> fake)0.6670.558
pp -value1.452988×10−101.452988\times 10^{-10}1.203812×10−21.203812\times 10^{-2}
Appendix
Table 16: Hypothesis 1 statistical results with ADE20K dataset, ResNet50V2 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Generation (Concept Alignment)Structure-Guided Fake Generation (Concept Alignment)Text-Guided Fake Generation (Activation-Based Relevancy)Structure-Guided Fake Generation (Activation-Based Relevancy)
Images paired888880269285
Nonzero diffs used888880269285
Median(real - fake)1.46780.10444.28300.8398
Mean(real - fake)1.37750.40814.32641.4776
Prop(real >> fake)0.7180.5360.9700.775
pp -value4.977×10−494.977\times 10^{-49}8.824×10−78.824\times 10^{-7}1.252×10−451.252\times 10^{-45}3.940×10−273.940\times 10^{-27}
Appendix
Table 17: Hypothesis 2 statistical results with ADE20K dataset, ResNet50V2 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used329215
Mean (real_count)0.4310.431
Mean (fake_count)0.0140.196
Median(real_count - fake_count)1.0001.000
Prop(real_count >> fake_count)1.0001.000
Appendix
Table 18: Hypothesis 3 statistical results with ADE20K dataset, ResNet50V2 CNN model, and Stable Diffusion image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used232205
Median (real-fake) counts1.0001.000
Prop(real >> fake)0.8280.546
pp -value1.044829×10−191.044829\times 10^{-19}1.176608×10−21.176608\times 10^{-2}
Effect size ( rrbr_{rb} )0.6550.093
Appendix
Table 19: Hypothesis 4 statistical results with ADE20K dataset, ResNet50V2 CNN model, and Stable Diffusion image generation model.
Figure 9: Suppression in neurons and their mean activation drop with ADE20K dataset, ResNet50V2 CNN model, and Stable Diffusion image generation model. (a) Text-guided fake image generation (all neurons), (b) Structure-guided fake image generation (all neurons), (c) Text-guided generation for confirmed neurons, and (d) Structure-guided generation for confirmed neurons.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used400329
Median(real - fake)1.0001.000
Mean(real - fake)0.8180.356
Prop(real >> fake)0.7180.626
pp -value2.304215×10−172.304215\times 10^{-17}3.436013×10−53.436013\times 10^{-5}
Appendix
Table 20: Hypothesis 1 statistical results with ADE20K dataset, ResNet50V2 CNN model, and FLUX image generation model.
StatisticText-Guided Fake Image Generation (Concept Alignment)Structure-Guided Fake Image Generation (Concept Alignment)Text-Guided Fake Image Generation (Activation-Based Relevancy)Structure-Guided Fake Image Generation (Activation-Based Relevancy)
Images paired925905316292
Nonzero diffs used925905316292
Median(real - fake)1.11810.16403.30380.8045
Mean(real - fake)1.04560.19933.47080.8583
Prop(real >> fake)0.6920.5720.9840.757
pp -value2.017×10−432.017\times 10^{-43}1.966×10−81.966\times 10^{-8}1.251×10−531.251\times 10^{-53}8.404×10−248.404\times 10^{-24}
Appendix
Table 21: Hypothesis 2 statistical results with ADE20K dataset, ResNet50V2 CNN model, and FLUX image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used321229
Mean (real_count)0.4310.431
Mean (fake_count)0.0210.191
Median(real_count - fake_count)1.0001.000
Prop(real_count >> fake_count)1.0001.000
Appendix
Table 22: Hypothesis 3 statistical results with ADE20K dataset, ResNet50V2 CNN model, and FLUX image generation model.
StatisticText-Guided Fake Image GenerationStructure-Guided Fake Image Generation
Images paired13701370
Nonzero diffs used256212
Median(real - fake)1.0001.000
Prop(real >> fake)0.7580.580
pp -value1.963471×10−141.963471\times 10^{-14}2.452926×10−22.452926\times 10^{-2}
Effect size ( rrbr_{rb} )0.5160.160
Appendix
Table 23: Hypothesis 4 statistical results with ADE20K dataset, ResNet50V2 CNN model, and FLUX image generation model.
Figure 10: Suppression in neurons and their mean activation drop with ADE20K dataset, ResNet50V2 CNN model, and FLUX image generation model. (a) Text-guided fake image generation (all neurons), (b) Structure-guided fake image generation (all neurons), (c) Text-guided generation for confirmed neurons, and (d) Structure-guided generation for confirmed neurons.
Figure 11: Additional Sample fake output images using SUN2012 and Stable Diffusion variants; (a): from Text-Guided Fake Image Generation; (b): from Structure-Guided Fake Image Generation.
Figure 12: Additional Sample fake output images using ADE20K and Stable Diffusion variants; (a): from Text-Guided Fake Image Generation; (b): from Structure-Guided Fake Image Generation.
Figure 13: Additional Sample fake output images using ADE20K and FLUX; (a): from Text-Guided Fake Image Generation; (b): from Structure-Guided Fake Image Generation.

Explore similar work

CardsList
Feb 23, 2026

SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images

The rapid advancement of generative models has made the detection of AI-generated images a critical challenge for both research and society. Recent works have shown that most state-of-the-art fake image detection methods overfit to their training data and catastrophically fail when evaluated on curated hard test sets with strong distribution shifts. In this work, we argue that it is more principled to learn a tight decision boundary around the real image distribution and treat the fake category as a sink class. To this end, we propose SimLBR, a simple and efficient framework for fake image detection using Latent Blending Regularization (LBR). Our method significantly improves cross-generator generalization, achieving up to +24.85% accuracy and +69.62% recall on the challenging Chameleon benchmark. SimLBR is also highly efficient, training orders of magnitude faster than existing approaches. Furthermore, we emphasize the need for reliability-oriented evaluation in fake image detection, introducing risk-adjusted metrics and worst-case estimates to better assess model robustness. All code and models will be released on HuggingFace and GitHub.

Aayush Dhakal, Subash Khanal, Srikumar Sastry +4

Washington University in St. Louis · Oak Ridge National Laboratory

AI-Generated Image DetectionOOD GeneralizationOne-Class Classification
AskPDF
May 20, 2026

Comparative Evaluation of Deep Learning Models for Fake Image Detection

The growing sophistication of GAN-based image manipulation presents significant challenges for digital forensics. This study compares the performance of four pretrained CNN architectures including VGG16, ResNet50, EfficientNetB0, and XceptionNet for fake image detection using a unified preprocessing and training pipeline. A dataset of real and manipulated images was processed through resizing, normalization, and augmentation to address class imbalance and improve generalization. Models were evaluated using Accuracy, Precision, Recall, F1-score, and ROC-AUC. VGG16 achieved the highest accuracy at 91%, with XceptionNet, ResNet50, and EfficientNetB0 each reaching 90%. EfficientNetB0 showed stronger sensitivity to fake images but reduced reliability on real samples, reflecting imbalance-driven bias. Limitations include dataset imbalance, overfitting, and limited interpretability, which affect cross-domain robustness. The study provides a reproducible baseline and underscores the need for balanced datasets, advanced augmentation, and fairness-aware training to develop reliable fake image detection systems.

Akhitha Pakala, Mohammed Mahir Rahman, Shahzad Memon +1

University of East London, London E16 2RD, UK

Image Forgery DetectionDigital Image ForensicsConvolutional Neural NetworksAI-Generated Image Detection
AskPDF
Aug 27, 2026

Hierarchical Channel Stacking: A Structured Decision Framework for AI-Generated Image Detection

Many synthetic-image detectors produce accurate predictions but offer limited insight into how those decisions are formed. This paper introduces Hierarchical Channel Stacking (HCS), a compact framework for AI-generated image detection that converts intermediate CNN activations into a structured 60-dimensional representation organized across three progressively deeper backbone stages. HCS uses per-channel Level-1 classifiers and a Level-2 aggregator to produce image-level predictions while preserving explicit hierarchical structure for analysis. On a benchmark spanning GAN and diffusion generators, HCS achieves 86.7% accuracy and 86.7% macro-F1 on the held-out test set. Stage ablation shows that the full three-stage system outperforms reduced single-stage and two-stage variants, indicating that the hierarchy carries complementary predictive information. Stage-level contribution analysis further shows that, in the analyzed detector setting, fake GAN and fake diffusion images exhibit distinct stage-level contribution profiles. These results position HCS not simply as a compact detector, but as a structured framework for studying how synthetic-image detectors assemble evidence across representation levels.

Saifullah Shoaib, Akash Borigi, Rupendra Lekkala +5

Missouri University of Science & Technology, Rolla MO 65409, USA · University of Houston, Houston TX 77479, USA · Aug Cranium AI, Inc., Short Hills NJ 07078, USA

Hierarchical ClassificationExplainable Image ClassificationAI-Generated Image Detection
AskPDF