For you, Facebook may be a virtual campfire where you can meet up with family and friends to share stories. However, the social network is also a digital wild west full of fraudsters, networks of fraudulent accounts and direct bullies.
The company has spent more than a handful of dollars to sum up these bad guys. It's hard to win the fights fast enough, and bad actors can change tactics when their old tricks stop working. Facebook is now turning to a simulated platform to tame it.
For more of it
Subscribe to the CNET Now newsletter to receive the most important stories of the day from our editors.
The social network introduced an AI-controlled system on Thursday that simulates user behavior via the real Facebook system. The idea behind the "Web Enabled Simulation" research project is to try out several possible approaches for dealing with harmful behavior at the same time. Facebook will work with academic researchers to test the program and investigate possible answers to problems. The goal is to find the most effective choice with the least disadvantages.
With this technology, researchers can observe simulated Facebook activities on a prototype that the company has called WW. Researchers can use various techniques, including machine learning, to collect data on WW. Isolated from Facebook, which most of us see, users are represented by bots that operate with AI algorithms. Some bots behave like normal people, but others do all kinds of nefarious activities, such as searching for innocent users to cheat. This shows researchers where vulnerabilities are, and allows them to introduce new restrictions on user behavior and see how the bad bots respond.
"For example, we can create realistic AI bots that want to buy items that are not allowed on our platform, such as weapons or drugs," said Mark Harman, a Facebook researcher, in a blog post that shows the web enabled simulation- Program was presented. "Because the bot operates in the current production version of Facebook, it can do searches, visit pages, send messages, and do other actions just like a real person could."
It's not difficult to say that Facebook hopes to learn how to correct its platform by studying its own western world.
The bots are unlikely to become sensitive, rise, and cause a moral dilemma that spans several years. But Harman called the research environment "Westworld"in an interview. (Westworld is not the public name of the platform.) The analogy makes sense because the platform allows researchers to shorten the time and try approaches that may not work without real users to deal with the consequences.
Harman said the WW simulator is versatile in the types of problems it can model. For example, fraudsters have predictable behavior patterns compared to other activities, such as fake accounts that pretend to be people in the United States, but are actually controlled from Nigeria or Ghana on instructions from a Russian intelligence agency. (And you thought the second season of Westworld was complex.) Harman says WW can model any situation.
WW can also deal with subjective situations, e.g. B. Content marked for abuse of Facebook's Community Guidelines. Human moderators in stressful, poorly paid work situations are currently dealing with these complaints. Harman said WW cannot apply subjective standards on its own. However, it can examine how moderators have dealt with complaints in the past and learn how to apply the standards in general.
The end result will still be "an approximation," Harman said, which means that people are needed to ensure that the judgments are correct.
WW relies on three techniques to simulate user behavior on the platform. The most coarse is called a rule-based algorithm, which defines if-then situations to control bot behavior. Here researchers work with what they already know about harmful user behavior patterns.
More advanced simulation techniques are supervised and unattended machine learning, which can also use information about defined behavior patterns. However, they also let the researchers learn to anticipate new behaviors that bad actors could try across the board. For example, researchers could give a bot a goal: to get other bots to fall for a scam. The bot doesn't have to follow strict rules to achieve the goal, which can reveal weaknesses that real bad guys haven't thought of yet.
The ultimate goal, said Harman, "is to find a mechanism that frustrates a real user who has a similar intention."
If WW works, Facebook will be faster at the draw if a new villain drives into town.
