Published 49 minutes ago

New AI Forensics Tool Maps Deepfakes Back to the Generative Model Behind Them

Detecting a deepfake is no longer enough. The harder problem—and the more consequential one—is knowing which AI system made it.

That distinction matters more than it might seem. If a fake video circulates online, knowing it’s synthetic tells you something. Knowing it was generated by a specific model, version, and development team tells you who to hold accountable, which systems are being exploited, and where to apply pressure to reduce future harm.

That’s the gap UC Riverside’s SAGA framework is designed to close.

4 mins read

6 sections

Key Highlights

SAGA shifts deepfake analysis from simple detection to full source model attribution
Temporal signatures act as involuntary fingerprints for generative video models
Designed for real-world pipelines, SAGA targets accountability and safer AI systems

From Detection to Attribution

The Source Attribution of Generative AI (SAGA) framework was developed by researchers at UC Riverside, including doctoral candidate Rohit Kundu, who also works as a research intern at YouTube. The project evolved through three distinct phases: detecting whether a video is AI-generated, explaining why it appears fake, and finally identifying the source model responsible.

That last step—source attribution—is where SAGA distinguishes itself from conventional deepfake detection tools. Rather than returning a binary “real or fake” verdict, SAGA is positioned to identify the specific generative model used, its version, and the team behind it.

The practical implication is significant. If a particular model is repeatedly linked to harmful synthetic content, that information can be shared with the model’s developers, prompting them to tighten filters, restrict certain prompt types, or patch the vulnerabilities being exploited.

The Mechanism: Temporal Signatures

The key technical insight behind SAGA is that different generative models leave different traces in how video frames evolve over time. The researchers call these temporal signatures (T-Sigs).

Even when two models receive the same prompt and produce visually similar output, the way individual frames transition and change over time differs in subtle but measurable ways. These inconsistencies are rarely visible to the human eye—Kundu himself has noted that he can no longer reliably distinguish real from AI-generated video—but they are detectable through systematic analysis.

T-Sigs serve as a kind of involuntary fingerprint. The generative model doesn’t intend to leave one, but the architecture of how it produces motion and temporal continuity reveals its identity regardless.

Built to Work Outside the Lab

A forensics tool that performs well on controlled datasets but fails in real-world conditions is of limited use. The SAGA team anticipated this and built the framework on top of a foundation model backbone—the core infrastructure that supports generalization across varied inputs.

When SAGA encounters video that differs from its training distribution, as will routinely happen when applied to social media content, the foundation model layer prevents performance degradation from domain shifts. This is a deliberate design choice, not an afterthought.

The framework also uses a video transformer architecture tailored specifically for attribution tasks. According to the researchers, this architecture is designed for practical adoption, meaning it can be integrated into existing forensics or content moderation pipelines without requiring a complete rebuild of surrounding systems.

The Broader Use Case: Accountability at Scale

The scenario that motivates much of this research is not abstract. Fake remote workers using deepfaked video calls have already caused real organizational harm. Synthetic video has been used in social engineering, identity impersonation, and disinformation campaigns. The volume of AI-generated video content on social platforms continues to increase.

SAGA’s attribution capability addresses a structural gap in how the industry currently responds to this problem. Most detection tools identify harmful content after the fact. SAGA adds a layer that connects that content back to its origin—which models are most frequently weaponized, what types of prompts generate harmful output, and which developers need to act.

The researchers frame this as a collaborative mechanism rather than a punitive one. The goal is to create a feedback loop: flag the source, inform the developer, improve the safeguards.

What Comes Next

The SAGA team has indicated that the next phase of research will focus on proactive prevention—stopping unsafe content from being generated in the first place, rather than identifying it after it has already spread.

That is a considerably harder problem. But the progression from detection to reasoning to attribution to prevention represents a coherent research trajectory, and SAGA’s temporal signature framework appears to provide a technically grounded foundation for the work ahead.

For teams working in content moderation, digital forensics, or enterprise security, the practical takeaway is this: the forensics tooling around AI-generated video is maturing faster than most organizations’ policies for responding to it. SAGA is not yet a commercial product, but the research direction it represents—model-level attribution, not just content-level detection—is where serious forensics capability is heading.