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

• Reliable detection remains important because AI-generated imagery is increasingly difficult to distinguish from photographs across GAN and diffusion generators.Source: §1 · p. 2
• Many accurate synthetic-image detectors provide limited insight into what drives their predictions.Source: §1 · p. 2
• 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.Source: §1 · p. 2
• 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.Source: §1 · p. 2
• 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.Source: §1 · p. 2

• Synthetic-image detection research includes CNNDetection and the GenImage benchmark spanning multiple diffusion generators.Source: §2 · p. 2
• Structured detector research explores alternatives to monolithic end-to-end classification, including frequency-based cues, transformation-based perspectives, and intermediate CNN representations.Source: §2 · pp. 2–3
• HCS differs by preserving explicit stage structure in its final representation, enabling direct study of how hierarchical evidence is distributed across source families.Source: §2 · p. 3
• Recent work emphasizes understanding detector behavior under changing generator distributions, alongside broader concerns about performance, robustness, and generalization.Source: §2 · p. 3

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.Source: §3.1 · p. 3
• 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.Source: §3.1 · p. 3, §3.2 · p. 4
• 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.Source: §3.2 · p. 4

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.Source: §3.3 · p. 4
• 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.Source: §3.3 · p. 5

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.Source: §3.4 · p. 5

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.Source: §4.1 · p. 6

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.Source: §4.1 · p. 6
• 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.Source: §4.2 · p. 6

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.Source: §5.1 · p. 7
• HCS performs better on GAN images than diffusion images, reaching 89.3% macro-F1 on GANs and 83.9% macro-F1 on diffusion images.Source: §5.1 · p. 7

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.Source: §5.2 · p. 8
• 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.Source: §5.2 · p. 8

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.Source: §5.3 · pp. 9–10

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.Source: §5.3 · p. 10, §5.4 · pp. 10–11

• HCS organizes intermediate backbone evidence into a structured 60-dimensional representation for directly analyzing detection decisions.Source: §6 · p. 11
• The full three-stage HCS system outperformed reduced single-stage and two-stage variants, indicating complementary predictive information across representation levels.Source: §6 · p. 11
• Stage-level analysis showed distinct contribution profiles for fake GAN and fake diffusion images within the analyzed detector and benchmark setting.Source: §6 · p. 11
• HCS enables examination of how evidence is assembled across representation levels, how source-family behavior differs, and how failure modes emerge.Source: §6 · p. 11