🤖 AI

SAGA: Source Attribution of Generative AI Videos

The paper introduces SAGA, a novel framework that leverages a data-efficient pretrain-and-attribute strategy and Temporal Attention Signatures to achieve state-of-the-art, multi-granular source attribution of generative AI videos, identifying specific models and developers with minimal labeled data while providing interpretable forensic insights.

Original authors: Rohit Kundu, Vishal Mohanty, Hao Xiong, Shan Jia, Athula Balachandran, Amit K. Roy-Chowdhury

Published 2026-04-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rohit Kundu, Vishal Mohanty, Hao Xiong, Shan Jia, Athula Balachandran, Amit K. Roy-Chowdhury

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet is a giant, bustling city where everyone is creating videos. For a long time, the only question people asked was, "Is this video real, or is it a fake?" It was like a security guard at a club checking IDs: Real (let them in) or Fake (kick them out).

But now, AI has gotten so good at making videos that they look almost perfect. The problem isn't just knowing if a video is fake anymore; it's knowing who made it. Was it made by a specific AI company? A specific artist? Or a brand new, unknown AI tool?

This is where SAGA comes in. Think of SAGA as a super-detective that doesn't just catch the criminal; it identifies the specific criminal's fingerprint, their gang, their style, and even the specific tool they used to commit the crime.

Here is how SAGA works, broken down into simple concepts:

1. The Problem: Too Many "Fakes" to Count

Imagine you have a pile of 19 different types of forgeries. A simple security guard (old AI detectors) can tell you, "This is a forgery." But they can't tell you if it was forged by "Artist A" using "Pen X" or "Artist B" using "Pen Y."

As AI video generators (like Sora, Runway, Pika) multiply, we need a way to trace a video back to its exact source. This is crucial for:

Forensics: Figuring out who spread a fake news video.
Copyright: Knowing who owns the style or the model.
Safety: Stopping bad actors from using specific dangerous tools.

2. The Solution: SAGA's "Two-Stage" Training

SAGA is smart about how it learns. It doesn't try to learn everything from scratch, which would take forever and require millions of labeled examples. Instead, it uses a Two-Stage Strategy:

Stage 1: The "Real vs. Fake" Boot Camp.
First, SAGA trains on a massive amount of data just to learn the difference between a real human video and any AI video. It's like teaching a detective to spot a fake ID card in general. It learns to see the subtle "glitches" or "artifacts" that AI leaves behind.
Stage 2: The "Specific Suspect" Specialization.
Once the detective knows what a fake looks like, SAGA teaches it to distinguish which AI made it. Here is the magic trick: It only needs 0.5% of the data to learn this.
- Analogy: Imagine you have a library with 100,000 books. Usually, to learn the author of every book, you'd need to read them all. SAGA is like a genius who reads just 500 pages, figures out the pattern, and can then identify the author of any book in the library, even ones it's never seen before.

3. The Secret Weapon: "Temporal Attention Signatures" (T-Sigs)

How does SAGA know the difference between two AI models that look almost identical?

AI videos have a "heartbeat." When an AI generates a video, it doesn't just make one picture; it makes a sequence of pictures that move. Sometimes, the AI gets a little clumsy with how one frame moves to the next. It might flicker, warp, or move in a weird rhythm.

SAGA invented something called T-Sigs (Temporal Attention Signatures).

Analogy: Think of every AI generator as a musician. Even if they play the same song, they have a unique "style" or "fingerprint" in how they hit the drums or slide the guitar.
SAGA listens to the "rhythm" of the video frames. It creates a visual map (a signature) of how the AI moves from one frame to the next.
The Cool Part: Even if SAGA has never seen a specific AI before, it can look at the video's "rhythm," see a pattern it doesn't recognize, and say, "This isn't the AI I know, but it has a unique rhythm. It's a new suspect."

4. The Five Levels of Detective Work

SAGA is a multi-level detective. It can answer questions at different levels of detail:

Level 1 (BIN-L): Is this real or fake? (Yes/No)
Level 2 (TASK-L): Did it turn text into video, or an image into video?
Level 3 (SD-L): Which version of the "engine" was used? (e.g., Stable Diffusion 1.5 vs. 2.1)
Level 4 (TEAM-L): Which company or research team built it? (e.g., Google vs. OpenAI)
Level 5 (GEN-L): Which exact model made this? (The specific fingerprint)

5. Why This Matters

Before SAGA, if a new AI tool popped up tomorrow, old detectors would be useless until they were retrained with tons of new data. SAGA is different because:

It's Data-Efficient: It learns incredibly fast with very little data.
It's Explainable: It doesn't just give a number; it shows you why it thinks a video is fake (by showing you the "rhythm" signature).
It's Future-Proof: It can spot new, unseen AI tools because it understands the fundamental "glitches" of how AI thinks, not just memorizes specific examples.

In short: SAGA is the ultimate AI video detective. It doesn't just say "This is a lie." It says, "This lie was told by [Name], using [Tool], with [Specific Style], and here is the proof." This helps us keep the internet honest in an age where seeing is no longer believing.

1. Problem Statement

The rapid proliferation of generative AI (e.g., Sora, Runway, Pika) has led to hyper-realistic synthetic videos, creating significant risks for misinformation and misuse. Current defenses primarily focus on binary detection (Real vs. Fake), which is becoming insufficient as models evolve.

The Gap: There is a critical need to move beyond simple detection to Source Attribution—identifying the specific generative model, version, or development team responsible for a video.
Challenges: Video attribution is significantly harder than image attribution due to:
1. Temporal Dynamics: Videos contain unique temporal fingerprints and motion inconsistencies missed by static image analysis.
2. Model Diversity: Video pipelines involve complex stages (frame synthesis, motion models), creating a vast attribution space.
3. Compression: Video codecs introduce spatio-temporal artifacts that can obscure subtle generator-specific traces.
4. Data Scarcity: Fine-grained source labels are scarce compared to binary real/fake labels.

2. Methodology: The SAGA Framework

SAGA (Source Attribution of Generative AI videos) is a comprehensive framework designed for multi-granular source attribution using a data-efficient, two-stage training strategy.

A. Architecture: Video Transformer with Temporal Attention

Foundation: The model leverages a frozen, robust Vision Foundation Model (pretrained on web-scale image-text data) as a domain-agnostic visual encoder to extract frame-level features.
Video Transformer: A novel multi-headed self-attention transformer processes these features hierarchically:
1. Spatial Encoder: Captures relationships between spatial patches within individual frames.
2. Temporal Encoder: Processes the sequence of frame-level vectors using Multi-Head Self-Attention (MHSA) to capture inter-frame dependencies and temporal inconsistencies.
Output: The model outputs embeddings ( $\phi_k$ ) representing the video's source.

B. Training Strategy: Pretrain-and-Adapt

To address the scarcity of labeled source data, SAGA employs a two-stage protocol:

Stage 1 (Pretraining): A binary classifier (Real vs. Fake) is trained on abundant data using standard Cross-Entropy (CE) loss. This establishes a strong visual representation.
Stage 2 (Adaptation): The pre-trained model is adapted for fine-grained multi-class attribution using only 0.5% of labeled source data.
- Loss Function: Instead of relying solely on CE loss, SAGA introduces a Contrastive Objective with Hard Negative Mining (HNM).
- Mechanism: HNM forces the model to separate overlapping clusters in the embedding space by focusing on the most difficult negative samples (those closest to the anchor but of a different class). This is crucial for distinguishing highly similar generators where standard CE loss fails.

C. Interpretability: Temporal Attention Signatures (T-Sigs)

SAGA introduces Temporal Attention Signatures (T-Sigs) to explain why different generators are distinguishable.

Derivation: By averaging frame-to-frame attention scores across multiple videos from a specific source, SAGA generates a unique visual "fingerprint."
Insight: T-Sigs visualize the subtle, stable temporal artifacts (e.g., characteristic motion dynamics or frame-to-frame inconsistencies) that the model learns to leverage. This provides the first visual explanation for video attribution performance.

3. Multi-Granular Attribution Levels

SAGA operates across five distinct levels of granularity, providing richer forensic insights than single-level analysis:

BIN-L: Authenticity (Real vs. Synthetic).
TASK-L: Generation Task (Real vs. Text-to-Video vs. Image-to-Video).
SD-L: Base Model Version (e.g., distinguishing Stable Diffusion 1.4, 1.5, 2.1, XL).
TEAM-L: Development Team (e.g., Stability AI, OpenAI, Alibaba).
GEN-L: Precise Model ID (e.g., specific model versions like Sora, HotShot, Latte).

4. Key Contributions

First Large-Scale Framework: SAGA is the first comprehensive framework for multi-granular AI video source attribution, moving beyond binary detection.
Data Efficiency: It achieves state-of-the-art performance using only 0.5% of labeled source data per class, matching fully supervised performance (100% data) through its two-stage HNM strategy.
Novel Interpretability: The introduction of T-Sigs offers a visual method to understand the temporal artifacts distinguishing different generators, even for unseen models.
Robustness: The framework demonstrates strong generalization in cross-domain and cross-dataset scenarios.

5. Experimental Results

The authors evaluated SAGA on DeMamba (19 generators, 1M real videos) and DVF (8 generators) datasets.

Binary Detection (BIN-L): SAGA achieved 99.94% accuracy on DeMamba and 95.39% on the cross-dataset DVF benchmark, significantly outperforming existing SOTA detectors (e.g., TALL, F3Net, ViViT).
Fine-Grained Attribution (GEN-L):
- Using only 0.5% data with standard CE loss, accuracy was poor (~24%).
- Adding HNM boosted accuracy to 65.80%.
- The full 2-stage SAGA framework achieved 94.99% accuracy, nearly matching the 100% data setting (97.41%).
Other Levels: SAGA achieved high accuracy across TASK-L (98.20%), SD-L (98.49%), and TEAM-L (97.77%) using minimal data.
Generalization: The model successfully attributed unseen generators and generalized across different teams and backbone architectures.
t-SNE Analysis: Visualizations confirmed that HNM creates clear separation boundaries between overlapping generator clusters, whereas CE loss alone resulted in significant overlap.

6. Significance

Forensic & Regulatory Impact: SAGA provides crucial tools for digital forensics, intellectual property enforcement, and regulatory compliance by pinpointing the exact origin of synthetic content.
Open-Set Recognition: The ability to generate unique T-Sigs for unseen generators suggests SAGA can flag content from new, unknown models, a vital capability for real-world deployment.
Paradigm Shift: The paper shifts the focus from "Is this fake?" to "Who made this?", offering a scalable, interpretable, and data-efficient solution to the growing challenge of AI-generated video provenance.