Source-linked AI summary
Hierarchical Channel Stacking: A Structured Decision Framework for AI-Generated Image Detection
Saifullah Shoaib, Akash Borigi, Rupendra Lekkala, Amaury Lendasse, Edward Ratner, Sai Sowjanya Bhamidipati, Alexander Schlager, Peggy Lindner
TL;DR
Synthetic-image detectors can be accurate while offering limited insight into how their decisions are formed. HCS addresses this gap by organizing intermediate CNN evidence into a structured three-stage representation, and the full hierarchy captures complementary information while revealing distinct GAN and diffusion contribution profiles.
Problem
Many synthetic-image detectors provide limited insight into the evidence driving their predictions, despite the growing need to detect images from GAN and diffusion generators.
Method
HCS converts selected intermediate CNN activations from three backbone stages into Level-1 estimates and a structured 60-dimensional representation scored by a Level-2 classifier.
Results
The full three-stage HCS system outperformed reduced single-stage and two-stage variants, while fake GAN and fake diffusion images showed distinct stage-level contribution profiles.
Takeaways & Limitations
HCS provides a structured framework for studying how synthetic-image detectors assemble evidence across representation levels rather than functioning only as a compact detector.
Takeaways & Limitations
The evaluation uses seen generators, and family-level contribution differences may partly reflect different benchmark sources rather than generator-family effects.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper frames AI-generated image detection as both a classification and interpretability problem, introducing HCS to preserve analyzable hierarchical evidence in detector decisions.
- Reliable detection remains important because AI-generated imagery is increasingly difficult to distinguish from photographs across GAN and diffusion generators.
- Many accurate synthetic-image detectors provide limited insight into what drives their predictions.
- HCS converts selected activations from three intermediate backbone stages into Level-1 probability estimates and a structured 60-dimensional representation scored by a Level-2 classifier.
- The framework preserves explicit stage structure so individual images can be analyzed by which stages push predictions toward fake or real, and by how strongly.
- HCS is presented as a compact framework whose contributions include structured decision representation, empirically meaningful hierarchy, and stage-level analysis of GAN versus diffusion behavior.
2 Related Work
Related work spans synthetic-image detection, intermediate-feature designs, and explainability, while HCS is positioned specifically as a compact detector-analysis framework preserving stage structure.
- Synthetic-image detection research includes CNNDetection and the GenImage benchmark spanning multiple diffusion generators.
- Structured detector research explores alternatives to monolithic end-to-end classification, including frequency-based cues, transformation-based perspectives, and intermediate CNN representations.
- HCS differs by preserving explicit stage structure in its final representation, enabling direct study of how hierarchical evidence is distributed across source families.
- Recent work emphasizes understanding detector behavior under changing generator distributions, alongside broader concerns about performance, robustness, and generalization.
3 Method
HCS converts selected intermediate ResNet-50 activations into a structured 60-dimensional representation through per-channel Level-1 models and a Level-2 classifier. Its stage-organized design also supports explicit decomposition of image-level predictions into contributions from progressively deeper backbone stages.
- Hierarchical Channel Stacking: HCS extracts activations from three progressively deeper ResNet-50 stages and retains selected channels from each stage.The stages are s2, s3, and s4; each represents a different level of backbone representation.
- Hierarchical Channel Stacking: The Level-2 random forest maps the 60-dimensional embedding to a final fake-versus-real probability, with threshold 0.5 used for reporting labels.This forms the two-level HCS inference design.
- Hierarchical Channel Stacking: Each selected feature map is flattened and scored by a channel-specific Level-1 random forest to produce a probability.The 60 retained Level-1 probabilities are concatenated into the image embedding.
- Training and channel selection: The top 20 channels per stage are selected by stand-alone macro-F1, producing the compact 60-dimensional representation.Selection uses a class- and generator-balanced validation split.
- Training and channel selection: Out-of-fold Level-1 probabilities provide leakage-safe training features for the Level-2 classifier.Five stratified folds are used for training embeddings, while validation and test features come from tuned models trained on the full training split.
- Stage-level contribution analysis: Exact groupwise Shapley values quantify how the s2, s3, and s4 groups contribute relative to a baseline prediction.Positive values support the fake class, negative values oppose it, and the contributions additively reconstruct the prediction up to negligible numerical error.
4 Data and Evaluation Protocol
HCS is evaluated on a balanced benchmark combining GAN and diffusion-model resources, using stratified train, validation, and test splits. Because splits contain images from the same generators, the protocol evaluates held-out samples from seen generators rather than unseen-generator or unseen-family generalization.
- Benchmark construction: The benchmark combines the Foren-Synths GAN resource with the GenImage diffusion-model resource.It is balanced across real and fake classes and across GAN and diffusion families.
- Evaluation protocol: The data are split into train, val A, val B, and test sets in a 70/10/10/10 ratio, stratified by class and generator.Train fits Level-1 models, val A selects channels, val B tunes hyperparameters, and test is reserved for final evaluation.
- Evaluation protocol: The held-out-sample protocol contains disjoint images from the same generator set across train and test.Therefore, it does not establish leave-one-generator-out or leave-one-family-out generalization.
5 Results
HCS is evaluated as a compact detector and as a structured framework for testing stage complementarity and analyzing family-dependent evidence. On the held-out benchmark, the full three-stage representation performs strongly, while its structured interface trades some raw accuracy against stronger black-box baselines.
- HCS performance and baseline comparison: 86.7% accuracy and 86.7% macro-F1 show that HCS preserves substantial discriminative information in its compact 60-dimensional representation.The representation is much smaller than the original intermediate activation space.
- HCS performance and baseline comparison: HCS performs better on GAN images than diffusion images, reaching 89.3% macro-F1 on GANs and 83.9% macro-F1 on diffusion images.
- Why the hierarchy matters: The full three-stage system achieves 86.7% test macro-F1 and outperforms every single-stage and two-stage variant.The ablation supports the claim that the stages carry complementary predictive information.
- Why the hierarchy matters: The strongest single-stage variant is s3 at 80.3 macro-F1, while the strongest reduced pair, s2 + s4, reaches 84.7 macro-F1.Neither reduced configuration matches the full three-stage system.
- What the hierarchy reveals: GAN fake images receive strongest positive support from s4, whereas diffusion fake images receive strongest positive support from s2.Representative examples reflect the same family-level stage pattern.
- Limitations and future work: These family-level profiles are specific to the present HCS instantiation and benchmark composition, not evidence of universal GAN or diffusion stage semantics.The evaluation uses held-out samples from seen generators, and source differences may contribute to the observed separation.
6 Conclusion
HCS organizes intermediate backbone evidence into a structured 60-dimensional representation that makes synthetic-image detection directly analyzable. Its hierarchical structure captures complementary predictive information and reveals distinct stage-level contribution profiles for fake GAN and diffusion images.
- HCS organizes intermediate backbone evidence into a structured 60-dimensional representation for directly analyzing detection decisions.
- The full three-stage HCS system outperformed reduced single-stage and two-stage variants, indicating complementary predictive information across representation levels.
- Stage-level analysis showed distinct contribution profiles for fake GAN and fake diffusion images within the analyzed detector and benchmark setting.
- HCS enables examination of how evidence is assembled across representation levels, how source-family behavior differs, and how failure modes emerge.