New Facebook AI Doubles Detection Of Fake Accounts

Facebook AI Doubles Detection Of Fake Accounts

Facebook has been using a new type of artificial intelligence (AI) to identify and delete hundreds of millions of fake accounts, according to a report by VentureBeat.

In its most recent Community Standards Enforcement Report, the company said it has removed 3.2 billion fake accounts between April and September of this year, which is more than double the same period a year ago.

According to Facebook, the increase in numbers is due to a machine learning framework called deep entity classification (DEC). The company said that in the two years since its implementation, DEC has resulted in a 20 percent reduction in accounts classified as abusive. While the company employs simpler models for detection during sign up, DEC works best in “challenge cases,” according to Sara Khodeir, a Facebook software engineer.

DEC was developed to go beyond traditional approaches of detecting fake accounts, such as identifying features like age, number of friends or location – details that attackers eventually figured out how to game. Instead, DEC looks at “deep features,” as many as 20,000 for each account, as opposed to only dozens or hundreds.

“Over the past few years that DEC has been in production, we’ve seen a step reduction in the number of [abusive] accounts on the platform,” said Khodeir. “Even though attacker volumes increase, DEC catches them at pretty much the same volume.”

The DEC platform will look at a number of features and then use aggregation to categorize them both numerically and categorically. The technology is able to identify fake accounts with a higher level of accuracy than what was previously thought possible, at more than 95 percent.

Facebook also uses language-agnostic AI to identify fake accounts. The technology is trained on 93 languages and 30 dialect patterns, and is used with other programs to handle many language problems at once. The company also uses AI to identify things in videos that it considers harmful or inappropriate.