EvoKnow: Continual Knowledge Evolution for AI-Generated Image Detection

1 Introduction

• Emerging generators may contain forensic traces absent from fixed training domains, while replay is costly and shared-model adaptation can overwrite prior knowledge.The challenge is adapting from limited current-domain evidence without retaining historical generated images or losing earlier forensic knowledge.Source: §1 · pp. 1–2
• EvoKnow treats forensic evidence as partly transferable and partly generator-specific, preserving the former while incrementally acquiring complementary evidence for emerging generators.A single evolving model risks forgetting, whereas independent models redundantly model common evidence.Source: §1 · p. 2
• EvoKnow freezes a shared forensic basis and prior experts, adds isolated residual experts, and updates AIR in closed form from current-stage images and accumulated statistics.The protocol retains neither historical generated images nor their instance-level features; task-agnostic real images provide negative-class supervision.Source: §1 · p. 2
• 96.70% average accuracy on non-base GenImage generators and 94.48% accuracy on Chameleon with ten generated images per arriving generator; continual learning reaches 96.32% mean stage-wise accuracy and 4.32% average forgetting.The Chameleon result is reported without target-benchmark adaptation; continual evaluation uses a strict replay-free protocol.Source: §1 · p. 2

2.1 Cross-Generator AI-Generated Image Detection

• Recent cross-generator detectors exploit transferable forensic representations, including frequency artifacts, reconstruction discrepancies, local textures, foundation-model features, and distribution alignment.These methods improve detection on unseen generators by using evidence shared across generative models.Source: §2.1 · p. 3
• Static cross-generator evaluation assumes fixed source domains, whereas emerging generators may show characteristics the source representation insufficiently covers.EvoKnow complements static detection by incrementally expanding coverage with generator-relevant evidence from newly arriving domains.Source: §2.1 · p. 3

2.2 Continual AI-Generated Image Detection

• Continual detection adapts across sequentially arriving generators, but replay requires historical samples and shared-parameter updates can interfere with prior knowledge.Related approaches include replay, regularization, constrained optimization, and parameter-efficient adaptation.Source: §2.2 · p. 3
• Analytical continual learning and subspace routing use sufficient-statistic updates and isolated experts; EvoKnow applies analytical retrieval while freezing shared evidence and adding complementary experts.AIR updates from current-stage images and accumulated sufficient statistics without retaining historical generated images.Source: §2.2 · p. 3

3.1 Problem Setting

• The setting provides only a few generated images per arriving generator, excludes historical generated-image replay, and uses task-agnostic real images only as negative-class supervision.Only learned model parameters and aggregated router statistics are preserved across stages.Source: §3.1 · pp. 3–4

3.2 Shared Forensic Knowledge

• The frozen shared forensic basis and base expert form transferable base knowledge, which subsequent stages preserve while adding generator-relevant residual knowledge.The shared basis is learned by the frozen shared forensic learner, and the base expert is trained on offline base-domain data.Source: §3.1 · p. 3, §3.2 · p. 4

3 Method

• The overview combines global and local views: cross-attention aggregates local evidence, which is fused with the global score for real-versus-synthetic prediction.Within experts, LoRA adapts shared semantic evidence while GPB and LFB emphasize structural and frequency evidence, respectively.Source: Figure 2 · p. 4

3.3 Incremental Residual Knowledge Unit

• Each arriving domain receives an isolated residual expert for complementary evidence insufficiently captured by the frozen basis, with unit-specific adaptations for global and local branches.LoRA modules adapt both branches; GPB emphasizes structural evidence globally, while LFB adds fine-grained spectral evidence locally.Source: §3.3 · p. 5

3.4 Analytical Incremental Router

• AIR selects the base or incremental expert whose forensic evidence is most compatible with an input, using routing features derived from the frozen representation.It incorporates new retrieval targets by regularized least squares with recursive updates and retains sufficient statistics, not historical images or instance-level features.Source: §3.4 · p. 5

3.5 Unified Multi-view Objective and Three-phase Optimization

• The unified objective supervises global, aggregated-local, and fused predictions with classification, uniformity, weak patch-level, and cross-view consistency losses.The three phases learn the shared basis and base expert, train only each new expert while freezing prior components, then update AIR in closed form.Source: §3.5 · p. 5

4.1 Experimental Protocol

• Experiments cover GenImage, UniversalFakeDetect, and Chameleon; arriving generators usually provide ten images, and historical generated images and instance-level features are neither retained nor replayed.Detector comparisons follow their source-domain protocols and are contextual rather than strictly controlled same-source evaluations.Source: §4.1 · p. 6

4.2 Generalization of Shared Forensic Knowledge

• Before expansion, EvoKnow (Base), initialized only on GenImage SD v1.4, reaches 93.66% average accuracy on non-base GenImage generators, 95.98% on UniversalFakeDetect, and 93.92% on Chameleon.The UniversalFakeDetect average excludes the ProGAN reference domain.Source: §4.2 · p. 6, Table 1 · p. 7

4.3 Few-shot Incremental Knowledge Expansion and Transfer

• With ten generated images per arriving generator, non-base GenImage average accuracy rises from 93.66% to 96.70%, while UniversalFakeDetect Ave. rises from 95.98% to 97.88%.Chameleon accuracy also rises from 93.92% to 94.48% without target-benchmark adaptation.Source: §4.3 · p. 7, Table 1 · p. 7

4.4 Replay-free Continual Knowledge Evolution

• 96.32% mean stage-wise accuracy and 4.32% average forgetting rank EvoKnow best among non-Joint methods on the 14-stage benchmark.The benchmark spans GANs, deepfakes, and diffusion models; EvoKnow does not retain or replay historical generated images.Source: §4.4 · p. 8, Table 2 · p. 8

4.5 Ablation and Analysis

• Combining global and local views improves balanced accuracy from 89.18% to 93.54%, and cross-view consistency plus weak patch supervision raises the full model to 94.40%.The authors interpret global structural context and local appearance evidence as complementary.Source: §4.5 · p. 8, Table 3 · p. 9
• Routing features raise linear-probe accuracy from 68.58% to 79.42% and k-NN accuracy from 67.83% to 78.83%, while related generators exhibit similar expert-selection patterns.AIR is used for expert retrieval rather than exact source attribution.Source: §4.5 · p. 8, Table 3 · p. 9, Figure 3 · p. 9
• Larger adaptation budgets improve UniversalFakeDetect average accuracy and reduce across-generator variation; ten images already improve coverage beyond the frozen shared basis.Additional samples mainly benefit generators that are challenging under zero-shot transfer.Source: §4.5 · p. 8, Figure 3 · p. 9

• EvoKnow preserves a shared forensic basis and incrementally adds isolated residual experts for complementary generator-relevant evidence.AIR retrieves expertise through accumulated sufficient statistics.Source: §5 · p. 9
• The framework expands detection coverage without retaining or replaying historical generated images or their instance-level features.Source: §5 · p. 9
• On the 14-stage benchmark, EvoKnow achieves the highest reported mean stage-wise accuracy and lowest reported average forgetting among non-Joint methods.Source: §5 · p. 9

A.1 Detailed Definitions of Auxiliary Multi-view Losses

• Patch-response uniformity forms spatial response distributions from patch logits and applies only to generated images, contributing zero when none are present.The target is a uniform distribution over patches, with temperature set to 1 in the experiments.Source: §A.1 · pp. 11–12
• Weak patch-level supervision averages global and local patch logits and applies image-level real/fake labels without requiring patch-level annotations.Source: §A.1 · p. 12
• Cross-view consistency aligns global, aggregated-local, and individual local decisions, alongside image-level classification supervision.The auxiliary objectives encourage spatially distributed, discriminative, and cross-view-consistent forensic responses.Source: §A.1 · p. 12

A.2 Implementation and No-Historical-Generated-Image-Replay Protocol

• The implementation uses DINOv3-L/16 with global and local views, a frozen shared learner defining the routing cut, and rank-8 LoRA residual experts with GPB and LFB modules.The model uses 224 × 224 inputs, one global view, four local views, and attaches the side modules in three deeper blocks.Source: §A.2 · p. 12, Table 4 · p. 13
• Each arriving generator normally supplies ten generated images for training a new expert while the shared learner and prior experts remain frozen.At each stage, 1,000 task-agnostic real images provide negative-class supervision; historical generated images and instance-level routing features are discarded, retaining only aggregated router statistics and learned parameters.Source: §A.2 · p. 12

A.3 Global Pattern-aware Block: Differentiable Structural Comparison

• GPB compares each patch's bottleneck features with those of its eight spatial neighbors and replaces the hard threshold comparison with a sigmoid relaxation.The learnable sharpness controls the transition; larger values concentrate sensitivity near zero feature difference, while smaller values spread it over a wider range.Source: §A.3 · p. 13
• GPB averages neighboring responses into local relative-order evidence, combines that response with bottleneck features, and adds a residual that directly modifies patch tokens.Reflection padding handles feature-map boundaries; class and storage tokens receive zero residual entries in the block.Source: §A.3 · p. 14

A.4 Local Frequency-aware Block: Learnable Spectral Selection

• LFB filters bottleneck feature channels in the Fourier domain and adds the transformed frequency residual to patch-token outputs.Its residual scale is initialized to 10^-3, and class and storage tokens receive zero residual entries.Source: §A.4 · p. 14
• LFB replaces a hard high-pass cutoff with a smooth radial mask, allowing the detection loss to provide gradients for learning the cutoff.The cutoff is constrained to (0, 0.5), initialized at 0.25, and shared across channels within each block.Source: §A.4 · p. 14, Figure 5 · p. 15

A.5 Analytical Incremental Router Details

• AIR initializes its base-expert router from base-domain generated images using routing features derived from the fixed shared representation.The routing features are comparable across stages because the shared learner and feature mapping remain fixed; the initialization uses ridge regularization.Source: §A.5 · pp. 15–16
• AIR expands its expert-selection outputs with current-stage generated samples, and its recursive update matches refitting the routing ridge objective under fixed features, regularization, and targets.The update avoids storing historical routing samples or gradient-based router optimization; AIR retrieves compatible experts rather than identifying exact generator sources.Source: §A.5 · p. 16

B Supplementary Experiments

• The supplementary experiments analyze complete continual trajectories and retention of historical knowledge.Source: B Supplementary Experiments · p. 16
• They study how the adaptation-sample budget affects cross-generator transfer.Source: B Supplementary Experiments · p. 16
• They also examine router persistent state, storage scaling, and auxiliary attribution behavior.Source: B Supplementary Experiments · p. 16

B.1 Continual Learning Trajectories and Knowledge Retention

• The GenImage sequence starts with SD v1.4, followed by BigGAN, VQDM, SD v1.5, Wukong, ADM, GLIDE, and Midjourney; the shared learner stays frozen after initialization.Seen averages accuracy over observed generators, while All includes all eight, including unobserved generators; the measures coincide at the final stage.Source: §B.1 · pp. 16–17, Table 5 · p. 17
• 94.41% at the first stage to 96.31% after the final stage: all-domain average accuracy improves overall despite fluctuations as generators of differing difficulty arrive.A decrease in Seen accuracy need not indicate forgetting because each new, potentially harder generator enters the average; when ADM arrives, Seen decreases while All increases and AF remains low.Source: Table 5 · p. 17, §B.1 · p. 17
• Low final forgetting accompanies residual-expert expansion without direct overwriting of shared knowledge or historical experts; router evolution can still change historical-domain performance.The section summarizes retention stage-wise and by generator at the final stage.Source: Table 6 · p. 17, Table 7 · p. 17, §B.1 · p. 18

B.2 Few-shot Adaptation and Cross-Generator Transfer

• Few-shot budgets vary from 10 to 300 generated images per arriving generator in the GenImage adaptation stream, with UniversalFakeDetect reserved for evaluation.The knowledge-unit organization and generator arrival sequence remain fixed across budgets.Source: §B.2 · p. 18

B.3 Incremental Residual Knowledge Organization

• The comparison holds knowledge-unit organization and arrival conditions to generator-based versus GAN, diffusion, and unknown-source category grouping.The default assigns one incremental residual expert to each arriving generator domain.Source: §B.3 · p. 18
• Category-based organization improves seven of the eight GenImage domains and raises the all-domain average by 1.79 percentage points, suggesting incrementally acquired residual knowledge is not entirely generator-exclusive.Source: §B.3 · p. 18
• The category-based result suggests related generators can benefit from partially shared complementary forensic evidence rather than entirely generator-exclusive residual knowledge.This is the authors’ interpretation of the seven-domain improvement and average gain.Source: §B.3 · p. 18

B.4 Analytical Router State, Resource Scaling, and Attribution Diagnostics

• EvoKnow is designed for binary real/fake detection and residual knowledge retrieval, not exact source attribution, which is reported only as a diagnostic.AIR selects residual units rather than explicitly predicting source identities; multiple generators may share related forensic differences.Source: §B.4 · p. 18
• “Replay-free” excludes retaining or replaying historical generated images, image features, and per-sample routing features, but does not mean zero persistent state.The router retains the regularized Gram inverse R, feature–target statistic Q, and weights W = RQ; checkpoints also include a fixed projection, unit mapping, and update metadata.Source: §B.4 · p. 18
• Router-state storage scales as O(m^2) for R and O(mM) each for Q and W; projection storage is O(md), and expert parameters grow as O(Mp).At m = 10,000, R occupies approximately 400 MB in FP32.Source: §B.4 · p. 18
• The nine-class attribution diagnostic has nonuniform generator-wise behavior, consistent with routing residual evidence for detection rather than enforcing a one-to-one generator-to-expert mapping.The evaluation includes one real-image class and eight generator classes; generator labels are supplied by the analytical router.Source: §B.4 · p. 19, Table 11 · p. 19