Anatomy of One of the Largest Deepfake Scams in its Day
Until recently, few details of the case were publicly available. Only in 2021 did it emerge when a request for asset seizure from the United States came before a Dubai court [1].
For the attack, the perpetrators used voice cloning technology and cloned the voice of the company director based on publicly available recordings: conference speeches, interviews, and YouTube videos [2].
This was a multi-layered attack. It was not a lone deepfake voice. Emails confirmed the authenticity of the information received. The fake lawyer sent documents. Everything appeared to be legitimate and routine.
17 People. 17 Accounts. 17 Countries. Not a Single Arrest.
Investigators found at least 17 participants in the crime. Funds were split between 17 different bank accounts located in different countries. Approximately $400,000 eventually reached US banks’ accounts, hence the involvement of the FBI [3].
Most funds were laundered via Hong Kong and onward to other countries. Up until this day, nobody was prosecuted, nor have the funds been recovered [4].
Why the Fraud Succeeded
The key to success was the banker’s familiarity with the voice of the company director. The fraudster did not use a stranger’s voice. He replicated the voice the banker knew very well. By 2020, deepfake voice generator technology allowed reproducing voice characteristics, including speaking speed, rhythm, intonation, and pauses.
To create a deepfake of a voice, scammers needed from 3 to 30 seconds of voice. With so many public voice records available, the creation of deepfake voice became possible for any executive.
This was the first recorded instance when deepfake voice generators were employed to commit an eight-digit fraud. Back in 2019, it was six digits; in 2020 – eight. This trend is clear.
How SYNHAWK Protects Against This Type of Attack
SYNHAWK PROTECTION: HAWK 7, SYNHAWK’s foundational model for audio, analyzes audio stream in real-time and recognizes AI deepfakes based on unique characteristics of AI speech. It recognizes artificial micro-compression artifacts, inconsistencies in phoneme transition, and signature artifacts of synthetic speech. Instead of relying on a fixed database of generators, it is designed to recognize new and unfamiliar deepfake generators.