The AI Girlfriend Scam: How We Detected a Deepfake Video Call

Background

A victim had been in a "long-distance relationship" for 6 months with someone they met on Tinder. They had regular video calls that appeared genuine. The "partner" then requested $15,000 for emergency medical expenses.

Investigation Methodology

1. Video Call Analysis: The victim recorded portions of video calls at our request. We analyzed the footage for deepfake indicators.
2. Facial Consistency Testing: Frame-by-frame analysis checked for micro-inconsistencies in facial geometry that are characteristic of real-time deepfake rendering.
3. Source Material Investigation: We attempted to identify the real person whose likeness was being used as a deepfake template.

Key Findings

• Video analysis revealed telltale deepfake artifacts: inconsistent facial boundary rendering during head movements, temporal flickering on the jawline, and unnatural eye reflection patterns.
• The deepfake template was traced to a South Korean model whose public photos were being used without consent.
• The scam operation was linked to a known organized crime group operating real-time deepfake romance scams at scale.

Outcome

The victim was prevented from sending the $15,000. The investigation report was shared with Interpol's cyber division and the platform's trust and safety team. Total investigation time: 1 week.