The report argues that the 3,000 percent increase in deepfake attempts and a fivefold increase in digitally forged identities in 2023 is due to the increasing availability of user-friendly generative AI tools. But businesses are using biometric verification powered by deep learning to deter deepfakes, combating AI-powered attacks with AI.
Less sophisticated varieties of fraud rose 7.4 percent from 2022 and made up 80.3 percent of all attacks and 79 percent of attacks against financial services, continuing a trend of bad actors going for quantity over quality of attacks to maximize reward.
Meanwhile, biometric fraud rates have doubled. Digital document forgeries rose by 18 percent, while physical counterfeits saw a 17 percent decline, though they still make up 73.2 percent of all document fraud.
The top documents targeted in document fraud are national IDs – accounting for 46.8 percent – followed by passports at 26.7 percent. National ID cards that are not for international travel don’t adhere to International Civil Aviation Organization (ICAO) guidelines and are therefore easier to forge.
Eighty percent of attempts to spoof liveness detection use videos displayed on a screen. Still, document-only checks have a 3x higher attack rate than biometrics, demonstrating that biometrics deter attacks.
In response to these findings, Onfido has released the Fraud Lab, which can mass-produce synthetic attacks. A release states that its Real Identity Platform can save customers over $3.9 billion in fraud losses. In six months, the Fraud Lab has improved fraud detection rates on documents by five times and biometrics by nine times. Fraud Lab experts measure fraud detection performance against attack vectors on document and biometric verification software.
Because it is challenging to train AI models off of a limited number of incidents, the team labels the key signals of the few samples they find and then uses the lab to generate synthetic 3D documents or deepfakes at a high enough volume to be useful as datasets for training and evaluation.