Retrieval-Driven Training-Free AI-Generated Video Attribution
AOCT-MSQR enables training-free video attribution by retrieval, reaching 84.6% Rank-1/78.3% mAP on GenVidBench at 100-shot.
Key Findings
Methodology
The paper reframes attribution as instance retrieval over a reference database rather than closed-set classification. Its pipeline applies Adapted Orthogonal Color Transformation (AOCT), Multi-Scale Quantized Residuals (MSQR), and weighted fusion of spatial residual, temporal-difference, and RGB streams. A Kinetics-400-pretrained R3D-18 extracts embeddings, and cosine similarity performs attribution or detection without task-specific retraining.
Key Results
- On GenVidBench, the 100-shot setting achieves 84.6% mean Rank-1 and 78.3% mAP for attribution; even with one reference sample per source, it reaches 32.8% and 52.0%, demonstrating useful few-shot fingerprint matching.
- For detection, 100-shot accuracy averages 91.0%, exceeding PiD at 90.2%, LOTA at 87.3%, and UNITE at 82.2%. Accuracy reaches 100% on T2V-Zero, ModelScope, and VideoCrafter2.
- Ablation raises mean Rank-1 from 51.1% for the base system to 62.0% with AOCT, 66.8% with MSQR, and 73.8% after adding temporal information, showing complementary spatial, quantization, and dynamic cues.
Significance
The work addresses a persistent forensic bottleneck: generators evolve faster than labeled datasets and supervised classifiers. Registering reference videos allows a new generator to enter the system without retraining, while retrieved exemplars provide more interpretable evidence than a bare class probability. The framework therefore extends video analysis from binary authenticity detection to source-level attribution, relevant to platforms, investigators, and regulators.
Technical Contribution
AOCT estimates a color basis from the principal eigenvector of a high-frequency residual covariance matrix, then uses simplex projection and Gram-Schmidt orthogonalization for a stable transform. MSQR applies channel-asymmetric quantization at multiple scales and reconstructs residuals through inverse transformation. Temporal residuals are first-order differences between residual frames. Fusion weights are 0.6, 0.3, and 0.1 for spatial, temporal, and RGB streams, respectively.
Novelty
The fundamental innovation is task reformulation: attribution is treated as source-fingerprint matching rather than classification. Compared with supervised systems such as VGMShield and SAGA, and image-oriented residual methods such as ESSP, LOTA, and PiD, the paper combines adaptive color analysis, multi-scale quantization, temporal aggregation, and retrieval into a training-free open-set video pipeline.
Limitations
- Evaluation is concentrated on eight generators and particular real-video sources in GenVidBench. Heavy compression, cropping, re-encoding, editing, or an entirely unregistered generator may weaken fingerprint similarity and cause incorrect attribution.
- Performance depends on reference-database coverage and the Kinetics-400-pretrained R3D-18 encoder. Dataset bias, generator-version drift, and semantic correlations can produce spurious matches; comprehensive compute-cost and cross-dataset analyses are not reported.
Future Work
Future work should test unseen generators and realistic platform distortions, add calibrated unknown-source rejection, and study confidence-aware retrieval. Stronger video encoders, local-region matching, compression-invariant fingerprints, approximate nearest-neighbor search, and auditable provenance databases could improve robustness and deployment scale.
AI Executive Summary
Highly realistic text-to-video systems are making visual inspection unreliable. Existing detectors usually answer only whether a clip is real or generated, while supervised attribution models depend on expensive labels and a closed list of generators. When a new model appears, they often require retraining. This paper instead treats attribution as retrieval: a query should match the reference video whose generator fingerprint is most similar.
The proposed Retrieval-Driven Training-Free AI-Generated Video Attribution pipeline first uses Adapted Orthogonal Color Transformation (AOCT) to expose informative channel variations from high-frequency residuals. Multi-Scale Quantized Residuals (MSQR) then amplify subtle discrepancies under several channel-asymmetric quantization scales. Spatial residuals, adjacent-frame temporal differences, and RGB context are fused with weights 0.6, 0.3, and 0.1, then encoded by a Kinetics-400-pretrained R3D-18. A new generator can be added by registering reference samples rather than retraining a classifier.
On GenVidBench, 100-shot attribution reaches 84.6% mean Rank-1 and 78.3% mAP; detection accuracy is 91.0%. With one reference sample, attribution remains 32.8%/52.0%. Ablations improve mean Rank-1 from 51.1% to 62.0% with AOCT, 66.8% with MSQR, and 73.8% after temporal aggregation. The approach is promising for scalable forensics, but compression, editing, unknown generators, and cross-dataset generalization remain open challenges.
Deep Analysis
Background
Video generators including T2V-Zero, ModelScope, VideoCrafter2, Pika, SVD, MuseV, Mora, and CogVideo increasingly approach natural-video realism. Detection research has exploited spatial, frequency, optical-flow, and long-range temporal cues, while VGMShield and SAGA address supervised attribution. Yet generated videos may preserve semantics while violating low-level statistical consistency, creating source-dependent traces that remain underused.
Core Problem
The goal is to identify the generating source under limited reference data, continuously changing generators, and simultaneous real-video detection. The task is difficult because fingerprints are subtle, scene semantics can dominate representation, and temporal dependencies introduce additional artifacts. Closed-set classifiers cannot naturally accommodate new sources, while single-frame image features miss flicker, motion inconsistency, and inter-frame dynamics.
Innovation
First, attribution is formulated as reference-database retrieval, enabling open-set expansion. Second, AOCT learns a high-frequency, covariance-based orthogonal color space rather than using fixed color coefficients. Third, MSQR uses multi-scale and channel-asymmetric quantization followed by inverse-transform residual reconstruction. Fourth, temporal residuals capture abnormal frame-to-frame fluctuations while RGB context preserves scene semantics. Together these form a training-free fingerprint representation.
Methodology
- �� Input: sample 64 consecutive frames and resize to 224×224; loop-pad shorter videos.
- �� AOCT: compute E=I−I*K with a 3×3 mean filter; form Σ=(1/N)E♭E♭ᵀ, take its maximum-eigenvalue vector, project it onto the constraints, and use Gram-Schmidt to construct M. Output Y=I×3M.
- �� MSQR: for scale S(i), compute Ŷ(i)=round(Y/S(i))S(i), then R(i)=I−M⁻¹Ŷ(i); average the residuals across scales.
- �� Aggregation: calculate ΔR_t=r_{t+1}−r_t and fuse F=0.6R+0.3ΔR+0.1I.
- �� Retrieval: encode F with R3D-18 and rank reference videos by cosine similarity for attribution or detection.
Experiments
GenVidBench combines real videos from HD-VG and Vript with eight synthetic sources: T2V-Zero, ModelScope, VideoCrafter2, Pika, SVD, MuseV, Mora, and CogVideo. Baselines include ResNet-50, R3D-18, ESSP, LOTA, PiD, DeMamba, and UNITE. Attribution uses Rank-1 and mAP; detection uses accuracy. The study evaluates 1-shot, 10-shot, and 100-shot settings, plus AOCT/MSQR/temporal ablations.
Results
At 100-shot, the method obtains 84.6% mean Rank-1, 78.3% mAP, and 91.0% detection accuracy; CogVideo attribution reaches 99.5%/95.0%. At 1-shot it still achieves 32.8%/52.0%. The ablation mean Rank-1 increases from 51.1% without fingerprint modules to 62.0% with AOCT, 66.8% with MSQR, and 73.8% after temporal fusion, confirming that spatial and dynamic evidence are complementary.
Applications
Platforms can register a small reference set for each known generator and simultaneously rank likely sources and detect generated content. Investigators can preserve retrieved exemplars and residual visualizations as supporting evidence. Deployment requires coverage of platform-specific compression, resolution, and generator versions, plus an unknown-source rejection policy to avoid forcing novel models into old categories.
Limitations & Outlook
Results come mainly from one benchmark and eight generators, so robustness across models, camera content, attacks, and real-world edits is not established. Although task-specific training is avoided, the method still relies on a pretrained R3D-18 and database-wide comparison; large repositories may increase latency. Future work should add approximate retrieval, perturbation augmentation, open-set calibration, multimodal provenance, and cross-dataset evaluation.
Plain Language Accessible to non-experts
Imagine several bakeries making cakes with the same design: a white plate topped with an orange squash. To a customer, the cakes look nearly identical, but each bakery’s machines leave tiny habits—slightly different edges, textures, color gradients, or occasional speckles. The paper does not train one strict judge to memorize a fixed list of bakeries. Instead, it keeps sample cakes from each bakery and compares a new cake with those examples.
First, the system views the cake through a special color lens that makes tiny marks easier to notice. It then uses several fine and coarse sieves to remove broad, uninformative appearance and preserve small details. Finally, it compares neighboring video frames, asking whether colors jump, edges shimmer, or textures change unnaturally. These details are combined with the overall scene, creating a kind of production fingerprint.
A new bakery can be added simply by supplying reference cakes. The same comparison can also flag cakes that resemble none of the registered bakeries, helping detect generated videos. However, if a cake is badly crushed, cut up, or produced by a completely unknown bakery, the evidence may be too weak for a reliable decision.
ELI14 Explained like you're 14
Suppose you are investigating who made a fake game clip. Eight apps can all generate a video of “a car speeding through a wilderness,” so judging by the story is almost useless. But every app has tiny habits: one makes outlines glow, another makes the sky flicker, and another leaves blocky patches. These clues are like software handwriting.
This paper builds a handwriting library. It breaks a video into frames, boosts details people normally miss, and checks whether nearby frames jump strangely. It also keeps the original scene, so the system does not confuse “car” with the app that made the clip. Then it compares the combined clues with stored examples and ranks the closest source.
The neat part is that the system is not locked to an old answer sheet. When a new app appears, you can add sample videos instead of retraining everything. With 100 reference examples per source, it gets the source right first 84.6% of the time on average, and detects generated videos with 91.0% accuracy. With only one example, first-choice attribution is still 32.8%.
But there is no magic! Heavy compression, extreme editing, or a totally unknown app can hide the clues. So this is best viewed as a smart investigator’s assistant, not an unquestionable judge. It narrows the possibilities and shows evidence for humans to check.
Glossary
Generative fingerprint
A stable, subtle statistical trace left by a generation model. It is not an explicit watermark; it emerges from the model’s synthesis process and can appear in spatial and temporal patterns.
The paper extracts fingerprints through residual representations and matches them by retrieval.
AOCT
Adapted Orthogonal Color Transformation, an adaptive color basis learned from high-frequency residual covariance. Orthogonality and channel constraints improve stability and perceptual consistency.
AOCT projects RGB frames into a space designed to expose generator artifacts.
MSQR
Multi-Scale Quantized Residual, which quantizes transformed frames at several scales and reconstructs differences through inverse transformation. It emphasizes fine-grained variations hidden by low-frequency content.
The implementation uses four quantization factors for the first channel and four for the remaining channels.
Temporal residual
The difference between neighboring residual frames, representing how artifacts change over time. It is sensitive to flicker, unstable gradients, and unnatural motion.
It forms the temporal stream in the final feature fusion.
Instance retrieval
A recognition strategy that ranks database examples by feature similarity rather than predicting a fixed class label. New categories can be added by registering examples.
The proposed system uses cosine similarity for both source attribution and detection.
Open Questions Unanswered questions from this research
- 1 Reliable rejection of unregistered generators remains unresolved; calibrated open-set thresholds and broader unseen-model testing are needed.
- 2 The effects of platform compression, cropping, filters, and repeated transcoding on fingerprints lack a systematic robustness curve and benchmark.
- 3 Large reference databases raise latency, privacy, and evidentiary-audit questions, especially in cross-platform deployment.
Applications
Immediate Applications
Platform content moderation
A platform can register a small reference set for each known generator, compute AOCT-MSQR features for uploads, and return both authenticity evidence and a ranked source list through cosine retrieval.
Digital-forensics triage
Investigators can preserve retrieved exemplars, similarity scores, and residual visualizations as supporting evidence. Final attribution should still combine these outputs with metadata, chain of custody, and expert examination.
Long-term Vision
Auditable provenance infrastructure
Generator versions, reference samples, timestamps, and retrieval evidence could form a shared provenance layer across platforms. Major obstacles include privacy, interoperability, adversarial manipulation, and legal standards for evidentiary use.
Abstract
AI-generated videos are becoming increasingly realistic and difficult to distinguish from authentic ones, which facilitates malicious misuse and poses growing threats to cybersecurity and social governance. Attributing AI-generated videos to their specific generative sources is therefore of critical importance for forensic investigation and legal regulation. However, most existing visual attribution methods focus on images and particularly rely on the image generation model, thereby lacking the ability to generalize to large-scale AI-generated video data. To address these limitations, we introduce an training-free AI-generated video attribution paradigm. Specifically, we formulates AI-generated video attribution as an instance retrieval task, and design a generative fingerprint-based pipeline. This pipeline consists of an adapted orthogonal color transformation, multi-scale quantized residual generation, and temporal-semantic aggregation, progressively capturing and integrating artifacts introduced by generative models across video frames. Extensive experiments on the GenVidBench benchmark demonstrate that our method achieves strong performance in both AI-generated video detection and attribution, outperforming existing state-of-the-art methods with a Rank-1 accuracy of 20.5% and a mean Average Precision of 16.6%. The code is at https://github.com/renxi-seu/Video_Attribution.