Let Them Fight

September 7, 2020 12:08 AM

With Valve unable or unwilling to combat cheaters in Team Fortress 2, an unofficial “Bot Extermination Service” has sprung up to deploy bots that hunt down and kill cheaters. The service joins others in the community fighting the scourge of cheating, including a player calling themselves “Anti Bot Bot” using hacks to target bots, and the TF2 Bot Detector tool that helps players vote-kick bots out of games.
posted by adrianhon (10 comments total)

How do they detect bots?
posted by Joe in Australia at 1:15 AM on September 7, 2020

> How do they detect bots?

"How does it work? It monitors the console output (saved to a log file) to get information about the game state. Invoking commands in the game is done via passing rcon commands to your client. Getting players in the current game is done via the tf_lobby_debug and status commands. Cheaters are identified by some rules but primarily by comparing players steamIDs against a list of known cheaters."
posted by genpfault at 1:25 AM on September 7, 2020 [1 favorite]

That's kind of disappointing actually. I was hoping that there was something distinct and identifiable about the way bots play. There almost certainly is, I think, e.g. the timing of commands and precision if movement, but I suppose that's not accessible remotely.
posted by Joe in Australia at 1:59 AM on September 7, 2020

Yes, this is a ML problem for the server owner (and this might mean we are essentially waiting for Moore's law to make the cycles cheap enough).
posted by jaduncan at 4:44 AM on September 7, 2020

I assume that once they're over-run by lizard-bots, they'll simply unleash wave after wave of Chinese needle-snake bots, and after that they've lined up fabulous type of gorilla-bot that thrives on snake-bots.
posted by Johnny Assay at 7:15 AM on September 7, 2020 [5 favorites]

this is a ML problem for the server owner

So, yes, absolutely, but...

Compute shouldn't be too much of a problem. You could run an inference say once a minute on a random set of keyboard logs, and it should be pretty cheap.

The two difficulties that come to mind are a) getting training data (maybe not sooooo bad if the anti cheat stuff already allows you to scrape arbitrary data from players), and dealing with variety in cheating programs. When you encounter a new program outside the training set, it's hard to say what the classifier will do. You also get the traditional abuse arms race, where the cheat programs will be actively trying to throw off the scent.

Machine learning is hard, adversarial machine learning is much harder. The nice thing about speech recognition is that the language is more or less the same from year to year, and doesn't actively change because you're doing a good job...
posted by kaibutsu at 9:06 AM on September 7, 2020

If you succeed in cowing AI opponents sufficiently they'll be undetectable, sure, but that will be because they're playing like humans. So you'll effectively have ended up with more people playing normally.
posted by Joe in Australia at 10:39 AM on September 7, 2020 [1 favorite]

Well obviously the plan here should be to use a GAN: design both a system to detect bots and to design undetectable bots in one go. It's basically the missing piece of skynet, so of course we should be deploying it on first person shooters.
posted by pwnguin at 12:19 PM on September 7, 2020 [3 favorites]

Game bot Turing tests are a pretty active research area.
posted by kleinsteradikaleminderheit at 11:28 PM on September 8, 2020

(That previous comment was apparently my 1000th comment on MeFi; I suppose it was inevitable that it would be a Simpsons reference.)
posted by Johnny Assay at 6:51 AM on September 10, 2020