SAGA: Source Attribution of Generative AI Videos¶
Conference: CVPR 2026
Paper: CVF Open Access
Code: None (Project page: https://rohit-kundu.github.io/SAGA)
Area: AI Security / Synthetic Video Forensics
Keywords: Video Source Attribution, Synthetic Video Forensics, Data-efficient, Contrastive Learning, Explainability
TL;DR¶
SAGA upgrades the question "is this video AI-generated?" to "which generator did it come from?". By utilizing frozen Vision Large Model features and a spatio-temporal dual-layer Transformer, combined with a two-stage strategy of "binary classification pre-training followed by contrastive adaptation with 0.5% labels," it achieves five-level source attribution from real/fake to specific models across 19 video generators. It further introduces Temporal Attention Signatures (T-Sig) to provide the first visual explanation of "why different generators are distinguishable."
Background & Motivation¶
Background: Faced with increasingly realistic AI-generated videos, mainstream defenses are almost entirely focused on binary "real/fake detection" (e.g., DeMamba, UNITE), which only answer whether a video is synthetic.
Limitations of Prior Work: As generative models explode in number, knowing a video is "fake" is insufficient. Digital forensics, IP accountability, and targeted mitigation require knowing "which model or team produced it." Source attribution has previously been performed almost exclusively on static images; the only existing work on video attribution (Vahdati et al.) covers only 4 closed-source generators and performs only single-granularity attribution.
Key Challenge: Video attribution is significantly harder than image attribution due to three hurdles: ① Temporal Fingerprints: Videos leave unique inter-frame motion artifacts during generation that static analysis cannot capture; ② Greater Model Diversity: Video generation involves multiple stages like frame synthesis and motion modeling, making the attribution space far larger than images; ③ Video Compression: Codecs introduce complex spatio-temporal artifacts that can mask or destroy the subtle traces of the generator. Furthermore, while binary labels are abundant, fine-grained source labels are extremely scarce.
Goal: To build a large-scale video source attribution framework that is deployable in real-world scenarios, data-efficient, and explainable, decomposing "attribution" into multiple granularities.
Key Insight: The authors observe that different generators leave stable and unique "fingerprints" in temporal attention, which remain distinguishable even for generators unseen during training. Consequently, temporal self-attention is used as both a discriminative feature and an explanatory tool.
Core Idea: Utilize "frozen VLM features + spatio-temporal dual-layer Transformer" to capture temporal artifacts, and a "binary pre-training → contrastive adaptation" paradigm to transfer knowledge from massive real/fake labels to multi-class attribution with only 0.5% source labels.
Method¶
Given a video \(x_k\), the goal is to predict its source label \(y_k\) from \(n_c\) candidates: \(n_c=2\) for binary classification and \(n_c>2\) for source attribution. SAGA's mechanism is not to train an \(n_c\)-class model from scratch, but to first pre-train a binary video Transformer on abundant real/fake data (Stage-1, using \(L_{CE}\) only), and then adapt it into a fine-grained attribution model (Stage-2, introducing a contrastive objective with only 0.5% source labels). Pre-training is performed once as a common starting point for all attribution granularities.
SAGA defines source attribution across five granularities (denoted as "-L"): BIN-L (Real/Synthetic), TASK-L (Real vs. T2V vs. I2V), SD-L (Distinguishing Stable Diffusion versions), TEAM-L (Attributing to development teams), and GEN-L (Specific generator ID). This hierarchy is crucial for practical forensics: if GEN-L gives low confidence for two similar generators, coarser levels like SD-L or TEAM-L can still provide valuable clues.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Video<br/>Frame-wise resize"] --> B["Frozen Vision Foundation Encoder<br/>Extract frame-level tokens zm, stack temporally into ζk"]
B --> C["Spatio-Temporal Dual-layer Video Transformer<br/>Intra-frame spatial attention + Inter-frame temporal attention"]
C -->|"Stage-1: Massive real/fake data + L_CE"| D["Binary Pre-trained Representation"]
D -->|"Stage-2: Only 0.5% source labels"| E["Hard Negative Mining Contrastive Objective<br/>L = λ·L_CE + (1-λ)·L_HNM"]
E --> F["Five-level Source Attribution<br/>BIN/TASK/SD/TEAM/GEN-L"]
C -.->|"Extract penultimate layer attention scores"| G["Temporal Attention Signature T-Sig<br/>Explainable Fingerprint"]
Overall Architecture¶
The input is a video, resized frame-by-frame and fed into a frozen vision foundation encoder (trained on web-scale image-text data to provide domain-agnostic features and mitigate the domain gap in real deployments). Each frame yields token embeddings \(z_m \in \mathbb{R}^{l_t \times d_t}\), stacked into a video-level representation \(\zeta_k \in \mathbb{R}^{L \times l_t \times d_t}\) (\(L\) is the number of frames). \(\zeta_k\) enters a trainable spatio-temporal dual-layer Transformer \(\theta\) to produce \(\phi_k = \theta(\zeta_k)\). Stage-1 uses a classification head \(\beta_1\) to map \(\phi_k\) to real/fake using \(L_{CE}\); Stage-2 adapts this pre-trained Transformer to \(n_c\)-class attribution with an additional hard negative mining contrastive loss \(L_{HNM}\). Simultaneously, T-Sig is extracted from the attention scores of the penultimate temporal encoder block as an explainable fingerprint.
Key Designs¶
1. Two-stage "Pre-train—Adapt" Paradigm: Leveraging real/fake labels for scarce source labels
This step addresses the scarcity of source labels. Instead of training a multi-class model directly, Stage-1 uses abundant real/fake data and \(L_{CE}\) to pre-train a strong binary video Transformer, establishing a general representation for "synthetic traces." Stage-2 adapts this base to \(n_c\)-class source attribution using only 0.5% of source labels per class. The benefit is data efficiency: on the GEN-L task, the two-stage scheme achieves 94.99% mean accuracy with 0.5% labels, nearing the fully supervised performance (97.41%) using 100% of labels (approx. 1.6 million samples).
2. Spatio-Temporal Hierarchical Video Transformer: Decoupling "Intra-frame Spatial" and "Inter-frame Temporal" Modeling
To address temporal fingerprints missed by static analysis, \(\theta\) processes \(\zeta_k\) hierarchically: first, Spatial Encoding refines the \(l_t\) tokens of each frame through a standard Transformer encoder block, followed by average pooling to obtain a single feature vector \(\mathbb{R}^{d_t}\) per frame. Then, Temporal Encoding adds sinusoidal positional embeddings to the \(L\) frame vectors to inject temporal order, passing them through \(D=\text{depth}+1\) stacked encoder blocks. Each block captures increasingly complex temporal dynamics/inconsistencies. The frozen foundation model provides domain-agnostic frame features, while the trainable Transformer focuses exclusively on cross-frame motion artifacts.
3. Hard Negative Mining (HNM) Contrastive Objective: Separating overlapping generator clusters
In GEN-L tasks, many generator embeddings overlap significantly in t-SNE space. The authors find that \(L_{CE}\) alone is insufficient, as it maximizes class separability in logit space but does not enforce geometric separation in embedding space. Standard semi-hard negative mining (semi-HNM) only selects negatives further than the positive but within a margin (\(\|a-p\|_2^2 < \|a-n\|_2^2 < \|a-p\|_2^2+\alpha\)). When clusters overlap heavily, many negatives are hard negatives (\(\|a-n\|_2^2 \le \|a-p\|_2^2\)) and are discarded by semi-HNM, resulting in insufficient gradient signals. HNM consistently selects the most difficult negative in the batch \(n_{\text{hard}} = \arg\min_{j:\,y_j \ne y_i}\|a_i-n_j\|_2^2\), with gradients \(\nabla_\theta L_{HNM} \propto 2(a-n_{\text{hard}}) - 2(a-p)\) that force the anchor away from the nearest heterogenous sample. The final loss is \(L = \lambda \cdot L_{CE} + (1-\lambda)\cdot L_{HNM}\). The impact is clear: Accuracy on GEN-L rises from 70.31% (CE+semi-HNM) to 94.99% (CE+HNM).
4. Temporal Attention Signature (T-Sig): Visualizing "Why It Is Distinguishable"
This is SAGA's contribution to explainability, answering why different generators can be distinguished. Inter-frame attention scores from the penultimate encoder block are averaged and normalized across many videos from the same source to obtain the T-Sig. These signatures capture stable yet subtle temporal artifacts (characteristic motion dynamics, inter-frame inconsistencies). Key observations: ① T-Sigs are highly consistent for videos of the same source despite varying content, while being visually distinct across different sources; ② Even completely unseen generators produce unique and recognizable T-Sigs, indicating the model learns fundamental temporal properties of synthesis rather than just memorizing training patterns.
Loss & Training¶
- Stage-1: \(L_{CE}\) only, pre-training the binary video Transformer on massive real/fake videos.
- Stage-2: \(L = \lambda \cdot L_{CE} + (1-\lambda)\cdot L_{HNM}\), adapting to \(n_c\) classes with 0.5% source labels; triplet margin constraint is \(\|a-p\|_2^2 + \alpha < \|a-n\|_2^2\).
- T-Sig is extracted during inference from attention scores of the penultimate temporal block, requiring no additional training objective.
Key Experimental Results¶
Datasets: Training on DeMamba (19 generators + 1M real videos), cross-dataset evaluation on DVF (8 generators). Three training regimes compared: [email protected]% data, 1-stage@100% data, and Ours [email protected]% data.
Main Results: Binary Classification (BIN-L) Cross-dataset Comparison¶
| Dataset | Method | Accuracy |
|---|---|---|
| DeMamba (In-domain) | MINTIME-CLIP-B | 89.98% |
| DeMamba (In-domain) | FTCN-CLIP-B | 89.67% |
| DeMamba (In-domain) | SAGA (BIN-L) | 99.94% |
| DVF (Cross-dataset) | DVF (In-domain training) | 92.00% |
| DVF (Cross-dataset) | HifiNet | 84.30% |
| DVF (Cross-dataset) | SAGA (Trained only on DeMamba) | 95.39% |
SAGA's performance on DVF is a pure cross-dataset evaluation (trained only on DeMamba), yet it outperforms SOTA methods trained specifically on DVF, demonstrating strong generalization.
Main Results: Multi-granularity Attribution¶
| Task Granularity | [email protected]% | 1-stage@100% | [email protected]% (Ours) |
|---|---|---|---|
| TASK-L (Overall) | 82.41% | 99.96% | 98.20% |
| SD-L (Overall) | 59.77% | 98.35% | 98.49% |
| TEAM-L (Overall) | 80.55% | 94.94% | 97.77% |
| GEN-L (Overall) | 24.55% | 97.41% | 94.99% |
With 0.5% data, 1-stage training collapses on difficult classes (e.g., SD 1.4/1.5 in SD-L drop to 0%, OpenAI in TEAM-L drops to near 0%, and GEN-L drops to 24.55%). The 2-stage paradigm recovers these almost entirely, nearing or even exceeding the 100% data setting.
Ablation Study: Impact of Loss Function on GEN-L¶
| Configuration (GEN-L, 0.5% data) | Overall Accuracy | Note |
|---|---|---|
| 1-stage, \(L_{CE}\) only | 24.55% | CE collapses under low data |
| 1-stage, \(L_{CE}+L_{HNM}\) | 65.80% | HNM significantly improves scores |
| 2-stage, \(L_{CE}\) only | 55.13% | 2-stage alone is insufficient |
| 2-stage, \(L_{CE}+L_{HNM}\) (Ours) | 94.99% | 2-stage + HNM synergy is optimal |
Comparison with semi-HNM: CE+semi-HNM achieves only 70.31% on GEN-L, while CE+HNM reaches 94.99%, confirming semi-HNM misses hard negatives in overlapping clusters.
Key Findings¶
- HNM is the decisive factor for fine-grained attribution: Finer granularities (GEN-L) depend on geometric separation in embedding space. HNM forces clusters apart by selecting the hardest negatives.
- Two-stage training provides low-data robustness: While single-stage fails on difficult classes, two-stage recovers class-wise performance, showing binary pre-trained representations transfer well.
- Coarse supervision facilitates fine-grained separation: t-SNE shows that even when trained on TASK-L/SD-L/TEAM-L labels, the model naturally clusters individual generators and makes unseen generators (like Hotshot, Show 1) separable, suggesting potential for open-set attribution.
Highlights & Insights¶
- "Five-level Granularity" is a pragmatic forensic design: When GEN-L is uncertain, retreating to TEAM-L or SD-L still provides useful clues, making it more robust than a single-output system.
- Multi-purpose Attention Scores: The same temporal self-attention serves as a discriminative feature and an averaged T-Sig fingerprint. This "discrimination as explanation" design is transferable to other forensics or anomaly detection tasks.
- Leveraging abundant labels for scarce ones: The two-stage paradigm essentially treats "real/fake detection" as a self-supervised pre-training for source attribution, a concept applicable to any hierarchical classification task with imbalanced label availability.
Limitations & Future Work¶
- Dependency on the feature quality of frozen vision encoders; sensitivity to encoder choice is not deeply discussed in the main text.
- Some specific generators remain difficult in GEN-L (e.g., DynamiCrafter at 56.64%, Sora at 73.33% in Stage-2); confusion between similar architectures is not fully resolved.
- Although compression robustness is mentioned as a motivation, systematic compression experiments are limited in the main text.
- T-Sig is currently a qualitative visualization with a lack of quantitative interpretability metrics; open-set recognition is observed in t-SNE but lacks a formal protocol.
Related Work & Insights¶
- vs. Binary Video Detection (DeMamba, UNITE): These only answer "real/fake," SAGA extends this to multi-granularity "which model/team," while outperforming them in binary detection cross-dataset.
- vs. Image Source Attribution (POSE, Wang et al.): These target static images and fail to handle video-specific temporal/motion artifacts; SAGA uses a dual-layer Transformer for inter-frame inconsistency.
- vs. Vahdati et al.: The latter only covers 4 closed generators at a single level; SAGA attribution covers 19 generators across five levels with explainable T-Sig analysis.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ First large-scale, multi-granularity AI video source attribution framework; T-Sig provides a novel visual explanation.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across 19 generators, 5 levels, in-domain/cross-dataset, and extensive loss/data ablations.
- Writing Quality: ⭐⭐⭐⭐ Motivation clearly outlined; Method description is solid, though many implementation details on compression are relegated to the supplement.
- Value: ⭐⭐⭐⭐⭐ Directly addresses real-world needs in generative AI governance and digital forensics with high data efficiency.