VeritasGaze: Forensic Deepfake Analysis for the Truth

About the Project

What It Is

VeritasGaze is a forensic deepfake analysis workbench that transforms deepfake detection from a binary “yes/no” classifier into an explainable evidence engine.

Instead of telling you “this video is 78% fake,” VeritasGaze shows you why — with frame-by-frame forensic proof that journalists, lawyers, and investigators can use in court or publications.

The system analyzes videos across three independent forensic domains:

Frequency Domain: Detects AI-generated artifacts via DCT/FFT decomposition
Temporal Consistency: Identifies lip-sync mismatches and unnatural mouth-audio correlation
Physics-Based Analysis: Finds lighting, shadow, and reflection inconsistencies

All results are aggregated into a confidence score with annotated evidence and exported as a court-ready PDF report.

The Inspiration: “Without Consent”

The Real Story That Changed Everything

We were inspired by Deutsche Telekom’s “ShareWithCare” campaign, especially the short film “Without Consent”, which follows a digital version of Ella — a teenager created from photos her parents shared online.

In the film, Ella says:

“All you need are a couple of pictures like the ones you share on social media where they can be taken and used by everybody... where my identity can be stolen just like that... where I can go to prison for things that I would never do... imagine my credit score being destroyed... or my voice copied to scam you…”

Watching this, we realized something critical: The video itself was a deepfake demonstration — used for good, to warn families about privacy. Yet, the same technology could easily be weaponized.

This paradox is the heart of VeritasGaze: If deepfakes can be created from innocent childhood photos shared online, society desperately needs tools to detect and prove them when used maliciously.

The Problem We Identified

The ShareWithCare campaign targets parents — encouraging digital literacy and careful data sharing. But it doesn’t address what happens when someone doesn’t listen.

By mid-2024, the deepfake landscape had changed drastically:

Anyone could generate convincing fakes using public tools
Detection systems turned into opaque black boxes

This created a massive gap:

Journalists got vague scores like “89% fake” with no explanation
Lawyers couldn’t use AI verdicts in court (inadmissible without reasoning)
Fact-checkers spent 6–8 hours per video manually inspecting fakes
Parents couldn’t prove their children’s likeness was stolen
Victims lacked evidence for law enforcement

In short, every commercial detector said “fake” — but none could prove it.

Our Realization

The Without Consent film showed a future problem. VeritasGaze delivers the present-day solution.

We chose to build not just another classifier, but a forensic workbench that answers:

“How can I prove this deepfake is fake — in a way that holds up in court, journalism, and public discourse?”

Our approach moves beyond binary detection to explainable forensic evidence, the same way crime labs analyze fingerprints or DNA — not just saying “match” or “no match.”