Facebook's DEC AI Detects Hundreds of Millions of Fake Accounts

2019-12-16

Facebook announced that they deleted over 3.2 billion fake accounts between April and September 2019, compared to over 1.5 billion during the same period last year. Part of this increase is attributed to a machine learning framework called Deep Entity Classification (DEC), which Facebook detailed for the first time during the 2019 Scale conference.


According to researcher Khodeir, it aims to address the issues Facebook encountered with its traditional automated fake account detection methods. Historically, the team would identify a set of features (such as account age, number of friends, and location) and label each feature as "abusive" or "benign" data for training the account classifier model. Because the features were handcrafted by engineers, the feature space was relatively small, making it easier for attackers to suspect. Eventually, these attackers started exploiting specific features of the system, such as using expired accounts to post harmful content.


In contrast, DEC extracts "deep features" of an account by aggregating behavioral attributes of other related accounts in the social graph. It is essentially recursive, with each account generating over 20,000 features, while only a few dozen or hundred are informative. It employs multi-stage, multi-task learning techniques that use a large amount of low-precision automatically generated labels and a small amount of high-precision manually provided labels, reducing the annotation work required before training.


DEC first considers the direct features of an account based on entity type, such as age and gender (user entity), number of followers and category (page), number of members (group), operating system (device), as well as country and reputation (IP address) spread across other entities that interact with that account. After feature extraction, aggregated entity aggregations are applied in a numerical way (e.g., average number of friend groups) and in a categorical way (e.g., percentage of most common category) before first and second-order filtering to produce results.This approach has been validated using three different models as well as a large amount of production data from Facebook: a behavior model with only direct features, a DEC model with tens of thousands of features, and a more complex DEC with a larger corpus. The results show that while the basic behavior model cannot predict fake accounts with over 95% accuracy, both DEC-based models surpass this threshold and identify more fake accounts.


Khodeir said, "In the past few years since DEC has been in production, we have seen a gradual decrease in the number of abusive accounts on the platform. Even as the number of attackers has increased, the catch rate of DEC has remained almost the same."


DEC is just one of the automated technologies actively used by Facebook to combat fake accounts and abusive behavior on its platform.

RECENT AI NEWS

RECENT AI TOOLS