GDC 2018: John McDonald (Valve) - Using Deep Learning to Combat Cheating in CSGO

I just watched a presentation about using deep learning to detect cheaters in CounterStrike: Go ( [VIDEO] ) and the question he didn't seem to have an answer for was data poisoning -- what if the cheaters all volunteer to be on the anti-cheater jury? Of course they are cross checking juror reliability ratings and stuff, but it's definitely a treadmill.

Valves John McDonald recently gave a talk at Gdc how they use machine learning to combat cheaters in CS:GO. Quite interesting to watch. [VIDEO]

If you wanna something more novel, look at this talk from Valve [0] on how they are integrating Deep Learning into their CS:GO cheater detection system (confusingly called Overwatch). It's not used to ban users, but rather to bring suspicious plays into their existing user-reviewed moderation system. In general, this type of human/AI side-by-side feedback loop seems to be very successful, all the way from games to moderating content on the web. [0] [VIDEO]

[VIDEO] Watch this! It's about using ML to catch cheaters. I searched youtube to show you this video; I used the search term "cs go machine learning". The first video was actually this: [LINK] Looks like it's using ML to make better aimbots--lol!

That's not true at all, though? Source has continued to get changes & upgrades over the years. It has a vulkan backend, the UI system is no longer scaleform but a web-based system called Panorama, a new physics engine called Rubikon that replaced the old Havok engine, etc... They aren't putting out many new games, although they did just release Artifact a couple months ago. But DOTA 2 & CSGO continue to get updates, changes, and improvements. CSGO for example has an entirely new, machine-learning based anti-cheat system called VacNet: [VIDEO] They have tons of talks on GDC of all sorts of really cool stuff they are doing that's related to Source 2 and has nothing to do with Steam.

> They also have a neural network setup that constantly scans match logs for cheaters For those wondering what this is referring to, it's a system called VACNet[0] presented at GDC last year. [0] [VIDEO] [1] [LINK]

Are you sure VAC doesn’t detect spin botters? Valve has implemented neural net based cheat detection trained from overwatch cases (players watching demos of reported players and judging if they were cheating). It’s a rather nice solution and doesn’t involve kernel mode drivers from a third party I don’t trust to make secure kernel mode drivers (a la battle eye or punk buster). [VIDEO]

You can't necessarily detect the cheat software, but you can detect the behaviour of cheat software. Writing an algorithm that can play a first-person shooter at a superhuman level is pretty trivial, but writing an algorithm that plays like a really skilled human is vastly more difficult. This behavioural approach to anti-cheat is now highly practical thanks to deep learning. [VIDEO]

Sort of reminds me of a talk Valve gave about creating an anti-cheat for Counter-Strike using AI. When asked if they were worried about people using AI to create cheats to fool the AI, his answer was essentially that it was an arms-race won by the person with more data/processing power. That person would most likely always be Valve. Link to talk: [VIDEO]

CS:GO also uses deep learning to automatically put cases into Overwatch, instead of relying on just user reports. Some information in this talk: [VIDEO]

See the video above. It’s a condensed recap of Valve’s GDC 2018 presentation on VACnet. The actual presentation is worth listening to if you have the time. [VIDEO] I disagree on the statement that developers, publishers, and/or anticheat companies are too incompetent to implement good solutions. They have not been putting serious effort into the problem or else the state of cheating in video games wouldn’t be as bad as it is today.

For those who don't know, Riot Games recently released a closed beta of Valorant, their new game with this anti-cheat. Valorant is a direct competitor to Valve's CS:GO, borrowing it's core loop and even some of its weapons. It's interesting to compare the two approaches to anti-cheat that the two companies are taking here: * Valve use VAC, an unobstrusive memory scanner in user-space with an emphasis on never having false positives. * Valve use machine learning on game replays to detect obvious aim-bots and spin-bots * Valve enlists the community via the "Overwatch" program to review game recordings and spot cheaters. * Valve supports CS:GO on Linux. * Leveraging your Steam history, Valve creates a "Trust Factor" score for every player and keeps trusted players together. * Riot uses ring-0 scanners. * Riot will not allow you to run Valorant in a VM. * Riot will ban specific hardware if a cheat is detected. Beware when buying 2nd hand motherboards! * Riot does not allow you to host your own Valorant servers (this makes hack-vs-hack servers, which some people enjoy, impossible) I think that in the long run Valve's approach, particularly with Trust Factor, will be proven the better of the two. [VIDEO]

Valve has an interesting talk about this with respect to VAC [1]. Which like you suggest relies heavily on statistical evidence from multiple replays in order to detect cheats. In CSGO's case they test it against their existing system Overwatch which uses player moderators to detect cheats). With their other big title dota as far as I'm aware its fully automated. 1 - "GDC 2018: John McDonald (Valve) - Using Deep Learning to Combat Cheating in CSGO" - [VIDEO]

Reminds me of the Overwatch moderation system used by Valve to collect cheater data from human evaluators in Counter-Strike. Eventually, they leveraged that data to improve their automated cheat detection systems. [0] Although Twitter's problem is way harder, IMO. [0] [VIDEO]

Exactly. Valve leads the way on non invasive anti cheat. These scummy anti cheat devs are not people I want with keys to my computer. Who is worth trusting? The answer is complicated and depends on many things. Anti cheat developers do not deserve anyone’s trust until they are open. [VIDEO] [LINK]

This presentation from Valve on VACNet touches on this: [VIDEO] Specfiically, at about 7:50 into this, he talks about the propensity of encountering cheaters in CS:GO. The tl;dr is that in a 5v5 multiplayer game the percentage of cheaters required to ruin 9 other players' games at once is not high. It only takes 7% of people to cheater to encounter 1 every 2nd game, and a 2% cheating rate gets you a 20% cheater-encounter rate. I'd also touch on the point that cheating ruins an online-only game. Like, it is the #1 way a game falls apart. Take a look at a game like Escape from Tarkov, where a cheater will ruin a good block of time for a lot of people just by existing in one game. It's especially bad in that game because the distinction between 'being caught by someone in a position you didn't check' and 'being killed by a cheater' is so unclear, and the consequences of losing are so high (and thus losing to a cheater is so much worse). It creates an environment where the players have to be absolutely 100% confident cheating doesn't exist, or it will very quickly sour people to the game.

But you can detect serverside (at least probabilistically) if a user is abusing it to their advantage. Some Valve guy gave a great talk about their cheating detection a while back; I found it incredibly impressive: [VIDEO] (can't comment on their effectiveness these days, haven't played CS in a long time)