Future of Anti-Cheat With Riot Games
Summary
- Client-side limitations: Traditional kernel-level anti-cheat becomes ineffective as cheaters advance to DMA exploits and AI-powered computer vision hacks.
- Server-side detection: Riot Games and Databricks shift anti-cheat strategy to server-side, data-driven approaches using machine learning and behavioral analytics.
- Vanguard insights: This video demonstrates real-world implementation of Riot Vanguard's adaptive security models detecting anomalies in player behavior and input patterns.
Future of Anti-Cheat With Riot Games

Talk by Carly Taylor, Field CTO, Gaming, Databricks ; Phillip Koskinas, Director of Anti-Cheat, Riot Games
As online gaming evolves, so do cheating methods that exploit client-server vulnerabilities. Traditional anti-cheat, such as kernel-level drivers and runtime detections, has long been the primary defense. However, advanced cheats like Direct Memory Access (DMA) exploits and AI-powered Computer Vision (CV) hacks increasingly render client-side detection ineffective. This presentation examines the escalating arms race between cheat creators and developers, highlighting client-side limitations. With CV cheats mimicking human behavior, anti-cheat must shift toward server-side, data-driven detection. By leveraging AI, machine learning, and behavioral analytics to analyze player patterns, input anomalies, and decision inconsistencies, future solutions can move beyond static detection to adaptive security models, ensuring fair play at scale. The session will also include real-life examples from Riot Games’ anti-cheat efforts, specifically insights and case studies from the development and operation of Riot Vanguard, to illustrate how these strategies are applied in practice.
Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements
Chapters
0:00Introduction and setup1:35Anti-cheat overview3:12Internal and kernel-level cheats11:29Trends in cheat vectors13:09Understanding pixelbots15:11Advanced 2PC pixelbot methods19:51Defensive strategies against pixelbots22:48Classification and behavior analysis27:20Behavioral detection models30:11Game state deception33:29Anti-cheat data architecture35:53DMA and hardware cheats38:20Hardware mitigation and conclusion
FAQs
How does AI detect cheating in online games?
AI systems analyze player behavior patterns, input anomalies, and decision inconsistencies to identify cheating that bypasses client-side detection. This video showcases how Databricks Data and AI platform enables machine learning models to detect sophisticated cheats like DMA exploits and computer vision hacks.
What are the limitations of traditional anti-cheat systems?
Client-side detection becomes ineffective against advanced cheats like Direct Memory Access (DMA) exploits and AI-powered pixelbots that mimic human behavior. This video explains how server-side detection overcomes these limitations.
What is Riot Vanguard?
Riot Vanguard is Riot Games' anti-cheat product for PC titles like Valorant and League of Legends. This video reveals real-world strategies and case studies from its development and operation.
How can machine learning improve gaming security?
Machine learning enables adaptive security models that continuously learn from player data to detect new cheating methods. This video demonstrates how Databricks Data and AI platform supports behavioral analytics for scalable, fair-play detection.
Full transcript
0:00 Introduction and setup
While we wait for people to filter in, I'm going to go through the most exciting part of the presentation. Uh there will be forward-looking statements in this presentation. will act like he knows what he's talking about. He doesn't. The most forward-looking. Uh, in some cases, forward-looking statements can be identified as using may, will, should, expects. Uh, forward-looking statements are based on what we know now.
Don't make any assumptions. Don't bet any money based on what we're saying and then come to us and get mad if you lose any. Uh, we're just up here. We don't know what we're doing. Complete your surveys in the app because everything has an app nowadays. So, get in there, do your surveys, tell them you liked us, tell them that I was better dressed than. Um, lie to her. Lie to me like everybody else. Uh, yes. Go in your attendee portal and do these.
I think someone actually reads them. So, you know, leave some jokes, leave some memes, make it fun. And we will get started right now, right on time, because we have a lot for you today. So, this is going to be maybe a bit of rapid fire. The way this is going to go is that Philillip is going to talk through what he knows best, anti-che. I am going to do my best to play the role of you, the audience, uh, and ask questions where I think that there might be things, unless you work in this industry, that you might not understand.
So, we're ready to go. You may now put on your 3D glasses. I thought you were going to change that subtitle. No, I left it. It was great. I thought they were going to make me take it out, actually, uh, because it doesn't fit the branding guides, but here we are. So, I've been talking. for those of you who know me or don't know me. Uh, I worked on Call of Duty for about five years on live ops the whole time. I still have the scars to prove it.
1:35 Anti-cheat overview
Uh, they're mostly mental. And I now make content on LinkedIn and I work at Data Bricks as a field CTO in gaming. Um, my best game, my favorite game is Jurassic World Evolution 2. Uh, three just got announced, which I'm so jazzed about. Uh, and my biggest fear is not providing stakeholder value. And uh Uh I'm the professor of anti-che at Vanguard Industries. Uh that is the anti-Geet product, the PC anti-Geet product that Riot Games develops to protect games like Valerant and League of Legends.
Well, most everything in this presentation will be about Valerant, but uh my best game is one called Super Hexagon. Um which you've probably never heard of, but I'm like number four on the leaderboard. Where you at? Damn. Let's afterwards maybe like Uh, and my biggest fear is male pattern baldness, to be honest, but, uh, I've dodged it so far. Um, all right. So, this is the menu. Um, we're going to talk a lot about the types of cheats that affect first-person shooters.
Mostly, I am going to use a lot of Valerant examples. Um, and you know, you can see it here, so I won't go through all of it, but we're going to be talking a lot about pixelbots and about um, the DMA cheating vector. Um, which, you know, we can go to the next slide. Uh yeah, so first up, encyclia. Um we're going to kind of describe those cheating vectors I just made mention of. And oh, before I get into all that, I'm going to use like vanguard terms to describe some of the stuff that we do.
3:12 Internal and kernel-level cheats
But it's it's a it's it's an active learning loop. Y'all are all data scientists, uh nerds really, so you probably know that we we're basically like utilizing the Vanguard product to query true false off of a client. Um that's how we kind of like create the detections. But from that we kind of want to predict the like we're trying to predict the actions that our agents take and um actually a ton of effort to predict the they're trying to basically predict bands the likelihood that any one person cheats and um yeah this is the core detection loop for the majority of anti-chets.
If you've ever worked on an anti-che it's probably not too dissimilar from those. I use Vanguard symbols to highlight that this is a Vanguard presentation. And if you are waiting for an architecture diagram for something like this, fear not. I have one for you later. So I see a lot of you getting uncomfortable waiting for that. Um, oh, quick note on the type of cheating. So, uh, and probably going to have something to say about this because she worked on COD.
Um, a lot of games that are built in Unreal and haven't opened like RPCs on the server or well really they have the ability to um they have the ability it gives the game studio the ability to architect a server that is authoritative. That is that it's simulating the majority of the game and the client is only really responsible for making requests to it. Um, this puts the developer in this position where they're kind of able to sanitize the majority of the exploits of old, like the things that you probably think of when you think of cheats, like god modes or infinite gold, infinite cooldowns, things like that.
When the server is responsible for simulating the majority of the game, you can outright reject the majority of this through either raw packet sanitization, parameter sanitization later in the function, or, you know, just Unreal actually does a lot of it for you. Um, what I will say here is that depending on the type of game you're working on, so specifically when I was working on battle royale games, there are times where you cannot be server authoritative.
I'm talking about times like when you're in a vehicle driving around a map. The only way for that to work to not feel laggy and terrible is for that to be client authoritative. meaning that all of the calculations for how your vehicle runs, the smoothing, the way that it handles terrain is all done locally on your PC or PlayStation or Xbox. So that means that is things like that are a vector to get hacked because anything coming from a client is inherently untrustworthy.
But there are moments where you are physically because of the laws of physics incapable of making things server authoritative. So if you're designing a game, be mindful there. It's that's true even in I'll extend that. It's true even in the first-person shooter genre. If you've ever played Valerant, there are characters within that game that can blink. Um, their blinks are actually deterministic. So, it can be server authoritative, but that blink and it's low time to kill of the game necessitate that the client is given a little more information than it needs.
And we'll get into a lot of that later in the presentation, but there are some there are some concessions you have to make in certain game design loops where your server can't be authoritative. Um but because uh we can go to the next slide. So um yeah so because every type the type of cheat describing before is something called internal cheating. So it's like basically you're trying to get code to run in the game clients or trying to read values.
You're trying to run something locally on the host for which the purpose is reading game memory or manipulating game memory to get the game client to send something to the server that it shouldn't do. That's like the typical exploit you're familiar with. But the majority of like we're talking about with the server authority that we're going to assume, we're going to baseline assume for the majority of the presentation.
Um then all you can really do through this type of cheating is gain access to something called an ESP. You're basically able to excfiltrate the um enemy positions. The enemies in in Valerance or in COD or in in Counterstrike too, the enemies uh you're able to get the positions for them and render them uh above the wall layer or on a separate radar. , what does ESP stand for? Oh, extra sensory perception. Wall hacks is uh fundamentally what they're trying to that's what's available in like a modern title.
But um it's it's not just uh it's not just reading. There are also writes that can allow you to do certain things even in the most secure titles. You could hook get direction, for example, or actually a lot of the way they do it in Unreal is they they hook a process event deep within the Unreal Engine and manipulate the parameters that way. Okay, it's everybody's favorite way to cheat in Unreal, but you're basically just updating your aim vector in game.
And the reason you might want to do that is to aimbot. However, like we're talking about before, modern anti-che has, for the most part, rendered internal cheating. Um, it's too detectable. The time to detection is too low. It's hard to productize for the cheat developer. It makes it difficult. It makes it's prohibitively difficult to get into when when you have I'm going to use the word invasive, but it's it's more about the fact that there's such a clean perimeter around the game client that you can't really read or write memory from Windows without alerting the antiche.
So, uh, as we'll get into this presentation, cheaters have pivoted into two other vectors, uh, that are are definitely our favorite. Um, one is pixelbots and one is DMA. And I'll do a little bit more on internal cheating if you take us to the next slide. Um this is kind of resummarizes everything I just said. Uh but as host security has improved to this stance where uh the perimeter around the game client is trusted we we basically know we can rely on wind especially on Windows 11 we can rely on a lot of hardware security features to know that the operating system is in a trusted state and if it is then we can trust that uh all the all the requests to read or write uh to valance memory are um are things that should be doing like things that should be doing that.
There are two ways that they typically get code these days into uh Valerant or any game client. And the majority and one of those is like just loading a signed library. And the reason that the reason it has to be signed is because, you know, with code integrity validation, we can outright reject uh loading any library that is not properly signed. And you're probably thinking, well, cheaters probably don't have signed libraries because they're operating out of uh some basement somewhere.
But you'd be surprised. uh they do get their hands on driver certificates for things like dishwashing service. Why it would need a device driver to instrument windows is beyond me, but that probably seems pretty obvious and it is. It's because obviously thatert has a has an entity. We can look it up and we can just b we can just deny list it, right? It's one of the reasons that vector is kind of not viable for something you want to sell because the minute 50 people have it, we're alerted to it.
We deny list. Um, so what you're telling me is that gamers hate kernel level antiche, but they would hate their lives more if it didn't exist. Oh, yeah. Easily. Well, especially for competitive games. I And I think that's I don't want to I don't want to spend a ton. We would be here forever if I justified all the reasons we need a kernel driver. But primarily, it's so we can do things like this second one, which is um Hey, sorry. the second one which is uh leveraging cheaters leveraging drivers of their own or leveraging a vulnerable driver to get their code into kernel memory or mapping it in manually.
There's a ton of ways they basically get code into Windows kernel. But when they do that um if we don't have kernel antiche then all the user mode all the calls we'd hook from user mode to try to alert ourselves to a read or a write occurring uh would be subtrifuge. They would be able to do it as if they were Windows itself. Um so uh like one of the things that we get with this on boot kernel driver is we have the ability to measure every driver that loads uh from Windows start measure it into the TPM and then through that we can kind of use it as this secure train of trust to know if if Windows itself was ever broken.
That is that one you know it's the who loads first problem. If their if their driver loads before us it can load in and then scrub itself from any telemetry that we might be able to obtain. So we load first and deliberately measure every every single driver that's loaded into the trusted platform module. I won't the reason we won't get into a lot of this is because it gets a little technically complex, but that's just the highle overview of what internal cheating is and now we'll get into the meat of the presentation.
It only took 10 minutes. Okay, we're right on time. Let's look at some graphs. Um I just said that. Uh you'll notice that the yaxis in this one has been elegantly scrubbed. Um, but what you're looking at is the number of uh matches. We call them infected matches. Basically, a match that has at least one cheater in it in Valerant. And you can see the you can see all the types of cheating on the right there in the legends.
11:29 Trends in cheat vectors
And um the three that we talked about before are the internal cheats which are rendered in the yellow. You can see there are a lot of them. Um and then also the pixelbots rendered in the green which we're going to get into and the and the uh the DMH which are red. very small subset of players. And you'll learn more about that in a minute. You're also probably thinking that, why would you use a stacked area graph? I have my eyes can't consume this or compare it in any way.
If only you had normalized them to each other or done something like a percent. Yeah. So anyway, this is a this is a stacked 100% graph of the same thing. Uh, and one of the things I want to highlight is that at like as I was talking about before, as host security improves, as we're able to outright detect and reject these internal cheats, we're seeing a greater percentage of both uh, cheat developers and the productized cheats themselves and the matches that are affected by pixelbots and DMA and DMA goes through waves and I'll get into that in a little bit, but um, we are seeing somewhat it's it's been seasonal and the seasonality is actually from when our anti-che is its strongest like when we're, you know, on top of our game and know all the cheats and then we kind of wayne and you can see that like the internal cheats come back and then we get strong again.
Um but we are seeing especially in the last three years an increase in the number of people that are looking at these external vectors Pixelbot and DMA cheaters seem pretty active in January. What are they up to? They don't have school. Oh yeah. Also we go on break and the timing could not be worse. So like most game studios have a winter break and cheaters it was always the worst coming back from that and realizing you know from experience I'm covered in scars.
13:09 Understanding pixelbots
Um, so Pixelbots, uh, at its at its base layer, there there I know I said I wouldn't talk about internal cheats, but there are both internal and external pixelbots. I'll explain what I mean. Um, but at its base layer, just imagine that your mouse is basically aiming for you. Um, uh, remember the client's basically secure. So they can't hook set direction within the game client like we had talked about before. They have to instrument their mouse directly.
Uh, and there are two ways they do that. So, why get good when your mouse can get good for you? Is that what we say? That's my favorite quote. Uh, imagine your mouse had a magstripe reader and you could just pay it to play the game for you. Wow, that sounds amazing. Why would you do that? You wouldn't. Um, but I have friends. Man, cold. Some cheaters have friends. They're just they also cheat. Um, all right. So, what you're looking at here is internal versus external uh pixelbots.
And I'll briefly kind of summarize. So in an internal pixelbot is something that is remember they're trying to manipulate the mouse to do so they have to read the screen for the opponent outline. Basically they're trying to update I I'll give a I'll give a pretty good example of it in a minute but they're they're trying to identify the opponent head or the opponent outline and then adjust the mouse to aim at it and then fire.
It's an aimbot, but it's an aimbot implemented totally outside of game memory. they're they're either using a screen classifier or they're just doing something like a simple color sampling algorithm to try to identify an opponent outline and then update the aim. If they're doing it internally, what this means is that the screen reader, the thing that's uh looking at the usually the center 300 by 300 uh pixels around your crosshair, they're that screen reader is local.
It's running on the host, which is good news because it means that we can probably find it. If it's external, and this is where it gets kind of crazy, they are doing it totally off the computer. And I'll explain more about that. So, so this is what um an internal tuh pixelbot looks like. Uh you basically have um your mouse going through a an Arduino or a KMB box. It's basically a uh you know it's a microprocessor for which the purpose is it is it is basically uh able to put instructions between what you did with your mouse as a human and what the game client eventually receives.
15:11 Advanced 2PC pixelbot methods
And the reason I've put a a little lock there is because Colonel Anti Jeets um the one that we develop has uh proprietary techniques to ensure that only one device is able to submit direct input to the game client. So they have they can't have like um they couldn't use virtual input with a script running locally. They ha it has to come from the mouse, right? So this is how they beat that. Now and in the internal t pixelbot architecture that Arduino is still connected only to one PC and it's being instrumented usually by a Python script that had is running like an OpenCV library.
Um but anyway next here's an example of a typical internal pixelbot. Um you can read that subtitle there because that's a little bit of a warning. Uh don't make me do it. I don't want to do this to you. It hurts me as much as it hurts you. Um every time I swing the hammer a piece of it swings back. Um, and it hurts. Uh, anyway, uh, here's the second thing I was getting into. The second the two PC pixelbot. Now, if you're running that Python script locally, like I was talking about before, we're going to notice that Python is reading the screen.
That OpenCV library, there's a ton of them that can be used to accomplish it, but eventually we're going to notice that the thing that's reading the screen doesn't have any right to be. And so, I shouldn't have triggerbot. py open while I'm playing. Well, no. Okay. I have to buy a whole It's funny she brings up so trigger bots are actually the most common type of internal cheat and I don't want to spend a whole lot of time on it but a trigger bot the difference between a pixelbot and a trigger bot is a trigger bot is only firing which is not as relevant in high time to kill games but in uh games like Counter-Strike or Valerant where you're you know you have uh an OP or an optical you literally have a scope on your sniper rifle and can look at head height across uh across a corner and literally the minute that a head enters it you'll fire.
Uh but anyway, advanced pixelbot is something that and it's this is the two PC cheating we're talking about is uh basically imagine a cheat that runs completely off of the host like we were talking about before. You still have that Arduino or that KMB box, but instead of being instrumented from the main PC, it's actually instrumented from a totally separate PC and you only need to get screen data undetected to that PC.
And there are three ways that people typically do that. Um one is leverage o OBS. This is actually the fastest one, curiously enough. The fastest way is actually like just to stream the 300 by 300 uh pixels at the center of your screen and then just uh have your second on the wire, your second PC just reads the MBD packets. Um but that's uh that's probably the fastest, but it's actually not the most popular. The most popular is a capture card.
So you just have a capture card on your and we don't know if you're using that because you're the most popular streamer in North America or because you know you just are recording all your gameplay. There are legitimate use cases for a capture card, right? So, we can't, and you'll notice that cheaters try to use things for which there is a legitimate use case because we can't deny it. So, we have to allow capture cards, but you could be using them to just send your screen data to a second PC.
And the third way is an HDMI splitter. This one's the least common, and it it's because they're expensive. Um, they add a lot of latency and the only ones that don't are basically another graphics unit, and they cost like $400 to $500, and it's just too large a price point for the people that want a colorbot. Oh, you also see me use Pixelbot and Colorbot interchangeably. did the same thing. Uh, and the final thing I'll say is once it's on that second PC, a lot of in Valerance, I I would like to pretend that there cheaters are smart enough to be doing screen classification or trying to do object detection to find enemy enemies to click on.
They're not. They basically just set their player outline because it's configurable for colorblind accessibility. They set the outline of the enemies to uh something outright like purple, pink, something bright that doesn't appear in the game otherwise. And then they color sample the the boxes in the center to find which one they this one has the most pink. I'll sample that one. And it's actually it's there's no classification whatsoever.
Um it is just a straight color sampling algorithm is trying to identify where an enemy outline is. How often do you fight with your devs about removing those outlines? Uh constant. Well, I've given up. So originally a lot. But uh they think it's and I think they're correct is that a lot of the accessibility of a tactical shooter is dependent upon the visual acuity of being able to identify an enemy. I know that I hated Counterstrike because of all the camo.
So I think that like I mean maybe I mean they got to be right. It's a I mean give it to them. It's a popular game. People love it or we wouldn't see so many cheaters. Um so yeah, like we were saying about before, once that second PC is uh in play, there's a lot of spare um there's a lot of spare graphical. You have another GPU to utilize, right? So they could use screen classifiers and they do in games like Call of Duty where the outlines aren't as bright or even in Fortnite where it's it's there's just so many colors it's not as applicable.
19:51 Defensive strategies against pixelbots
Um so they're actually are using simple simplified um object detection to try to find enemies to click on. But this is not the case in Valent. It is usually just color sampling. Um, and then you're probably asking, actually you're probably not asking. Maybe Sarah's asking, "Uh, why can't we just uh you notice that an Arduino is connected to the PC?" Uh, well, they do a thing called USB pass through, which is basically any query we could send the USB controller for that device.
They just take it and send it to the mouse that's connected to it and then reply with whatever the manufacturer intended for those things. They don't do it perfectly yet, but they're so damn close that we're worried about banning everybody off of the remaining detection surface because we would basic it would it would set us up for the inability to create ground truth of models we're trying to develop to address this stuff.
Um, and last thing I'll say is that like you saw that green graph before which was a comparison of external and internal pixel bots. It's about external pixelbots um are about 15% globally and it's growing by about half a percent a month. Uh that is they're completely external pixelbots. They're no longer reading screen on the host. Uh they're doing it by one of the three methods we talked about before. And uh that's got to be your Yeah.
Yeah. That's your I can always tell when it's one of your subtitles, but yeah. Um so how can we beat this stuff? Uh I kind of already brought up some queries we can send to the USB host controller that would try to make it reveal itself. But uh after that and after you've kind of secured direct input to only have one device and I it sucks to say this is an anti-cheater, but that's all the detectable surface you have from the host.
Everything else is scrutinizing raw input um or scrutinizing server sided aim vectors. Uh, and there's there are some prevention methods that we won't spend a lot of time on because I want to get to the rest of the presentation and leave a few seconds for questions, but there are things you can do like brought up before where you can manipulate to make it more difficult to colorbot. You can manipulate the uh the outlines in game.
Um, let's get into the first we'll get into the This one was you. Yeah, that one was me. So a lot of what I'm saying is that uh they're able to they're instrumenting their mouse externally and we're trying to basically determine that it's inhuman input, right? Uh this starts to become a war of algorithms, a war of a war of machines really begun the machine war has. Um and we're basically trying to identify uh that the raw input coming from the mouse or even the server sighted input uh the server sided aim vectors are uh inhuman.
We'll start with raw input classification because that's probably the one that has the highest granularity. I don't know what the tick rate is of every game server out there, but Valance is pretty fast. Um, and even with that 120 tick rate, we can't it's still it still takes us a lot of samples to actually know that a that a player is cheat is not human, that aim is not human. Whereas raw input, we can sample utilizing uh Vanguard, we can sample it at the exact same polling rate that the mouse is.
22:48 Classification and behavior analysis
Um, what you see on the left here is actually something called a a recoil bot. Um, it's basically like it's every time you click it's try it's it's just it's uh moving your mouse subtly downward. Um, trying to correct for the recoil from firing the rifle in the game. But one of the things that doing this kind of raw input classification necessitates is that you have like superb security posture because one thing I mentioned before is that they can actually just hook direct input in the game.
And if they do such then your your hook on raw input would receive nothing for the updates that the aimbot is making because it's instrumenting them directly in the game client. So just the and also kind of like if you just let's say you had a user mode hook for recording input, it could be instrumented by a a cheater will eventually find that and perform a sort of replay attack where they send a different set of data that's known human while they continue to cheat a game.
So you need you actually need a pretty secure setup to even get into this game. But once you are, you can start to sample um you can start to sample the it's actually just time series data for mouse input. Uh the the delta in XY every at the same frequency of the mouse. And um there's probably two modeling approaches we've tried. Uh there's obviously the the 2D uh convoluted neural network that you know uh nerds in the audience probably already saw this problem.
We're like I would love to throw that at it. It works. Uh it actually works very well. The issue is actually that you need ground truth. You need cheat data that you have to cook yourself. So you can either outsource a bunch of people and have them be cheating all the time or you could just have an anti-che that labels all the cheaters for you. Uh and it kind of closes that active learning loop we were talking about before and makes it really easy to identify um anomalous mouse stuff.
There's also vendors in the space and I'll give a shout out to Anyrain. is one of the ones that we've done a lot of research into and um they're probably the one that they're the most game agnostic one like they've trained it on everything they can. So their their synthetic moves model is best positioned to detect like a wide variety of um of pixelbots. Uh and I make one note here that if you are trying to do this if you are a game studio w don't interpolate or down sample the mouse of it.
You're going to be you're going to you're going to be encouraged to do this to reduce the size of the array you basically have to send back to your server. Um, but don't because it'll inevitably smooth the aimbot. There's no there's nothing you can do to kind of avoid this problem. But if you if you downsample the granularity, you'll lose visibility into the into the subtle adjustments that the aimbot is making because it's basically an aimbot arguing with a human.
So, you need a high sample rate. Next, one thing I will also say as talks about, you know, the issues that you have with sampling and the mouse and then your tick rate on your server, right, that you're losing information from the server basically not refreshing that quickly. Is that the same thing happens with your kill cams. So, if you're looking at a kill cam from a game to tell you what actually happened and whether or not you were hacked against, you're starting with a half deck of cards, right? that's not the reality of what the game what happened in the game world.
Um, and so you are also falling victim to the same thing where you're the data you're looking at is not complete. Really great point and I'll add that uh as games release replay systems that are relying only on what the server sees. If there's no lag correction, if there's no package drop correction, sometimes you'll see something we like to call a hackus, but it's basically players that see a snapped kill cam. And even if that uh the you know the client dropped a bunch of um a bunch of packets in between the update all the server saw was not pointing at head to pointing at head and it uh because so much of antiche is the perception of cheaters within the game.
Replay systems can actually become a pretty big enemy if you can't educate if you can't educate the player on um on on how they function. So now you all know and now you all have to go out and tell everyone else. We're relying on you to educate everyone else. Okay. We're going to check in in in a couple months and see if you've been evangelizing that kill cams will wreck your Okay. Uh so this is server sided classification which is not as good due to the granularity thing that we kind of already highlighted but it is really good at drawing a line around what's humanly possible and I mean human performance within the game.
27:20 Behavioral detection models
Um I have a member of my data team here with me today that probably recognizes these PR curves. Uh, but these are us attempting to basically uh like from just the the aggregated behavioral the aggregated in-game direction behavior trying to determine if the person is a human and it's it's imperfect. Obviously, we're not able to take immediate action. It takes I think we need like six games of you rage hacking to really get you out of game fast enough, but it's um that layer uh doesn't rely on um us excfiltrating all that mouse information and getting it through the inference.
It is really just looking at the server sighted aim samples and aggregating them in um well some of it's aggregate and then I there's a couple of things that are simple like you can just look at the maximum mouse velocity because anybody that's like clearly just running a rage hack pixelbot is going to snap all over the place instantaneously and you can kind of draw a very obvious line around what should be humanly possible within the game and anybody that goes outside of that you can get them out of the game faster.
The difficulty in this approach at that granularity with aggregated features is that a lot of pro players when they hop on their alternate accounts at lower MMRs, they massively overperform. Um, I mean, they're basically a rating a very high rated player playing against uh people who are totally new to the game and they just stomp them and they'll inevitably look like cheaters to models like this. Um, and that's something we're kind of and you can actually see version one which was like three months ago and then version two which was like last week.
So it is improving very quickly as we figure out like which features we need to add to improve our ra detection. So if you are building a game, one thing I will say that brings up is that the more people you have in your lobby, the higher likelihood you have a very great skill disparity in your lobby. So, if you have like pro players smurfing and they're in these lobbies where they're maybe they don't look like they're that good, but they're actually really good, the high skill disparities can really mess with your behavioral antiche because you're doing these aggregations on filters.
You're looking to see, you know, is this person performing above what we would expect for someone at their level or for someone in this lobby, right? You have an expected outcome for every lobby matchup. That's what matchmaking does, right? You can kind of predict who's going to place what in any game match if you know their skill ahead of time. put ranked rating as a feature in the model 100%. Okay. Yeah. Um, oh, I was going to make a joke earlier that I really regret not making, but on the raw input classification, we actually had a lunatic on our data team that just threw like ResNet on like literally on the graph you saw of the user's input and it performed as well as the time series model.
I mean, it is two dimensional data, right? So, it's it's a longer way to get to the same thing, but it was just what a lunatic. Um, so here's a kind of looking at those behavioral things we've created that are meant to address pixel bots. And as you can see, my analytics team is just getting lazy. Um, no, I'm kidding. The so the the bottom the darker blue is uh the manual bands. Basically, somebody that reviewed input, reviewed uh actions taken in a game and determined them to be cheating.
30:11 Game state deception
And the light blue is that slowly being totally replaced by the models. Uh they're coming for our jobs and it's all over. No, obviously we still have to uh a lot of the review takes longer now because these are things that have bypassed the model. They're performing within human means and um you know it's a it's an arms race with antiche. Uh so we're not getting as many obvious cheaters to review but we're finding new and new things and um oh this is your slide.
Oh yeah. So this is something that I actually built during my time at Activision with uh someone I was working with at. So, what we were trying to do was basically we're messing with all different kinds of cheese that we've talked about today, internal and external. But I think this goes well with what we've been discussing here, which is can you confuse a computer vision algorithm by putting things into the game that maybe shouldn't be there.
So, with this example, we called this hallucinations. So, we were actually fighting internal and external cheats. Like I said, from an internal perspective, the way that we injected these into the game client was actually pretty brilliant, if I say so myself. Um, we were per client that we suspected to be cheating, inserting a new, uh, player into the snapshot in memory getting sent from the server. Because if you do this, if you, you know, basically put a bot into the game and have cheaters say like, "Oh, if they shoot at a bot that's hidden behind a wall, they have a wall hack."
That's great. but they'll just actually look at the server snapshots. Delta, when you actually put this thing in and know it's not a real player. So, we had to have the match from their perspective start with this player that from looking at the snapshot you would assume was real. So, what we did was we just actually did a memory copy of another real player in the game. So, when you did this in production, it was funny.
You could just pick a random person like playing War Zone, do a mem copy of them, inject them into the snapshot of the cheater, and they would see this person doing whatever they were doing, but from a different part of the map. So, they'd be like trying to run through a door. So, it would just look like someone like not knowing where they were because they weren't actually in that part of the map. It was just a copy.
So, you can do this though to also confuse like a computer vision system. So, you know, maybe with the outlines, you could start to make the outlines a bit fuzzy. Do something if you're running a true computer vision. you know, they're pretty brittle to things that change. So, change the color up a little bit. Something that the human eye screen for one frame and see what happen. Yeah. If it shoots at everything, right?
Like do you could get really complex with this. So, you can spawn fake enemies within geographies. You can spawn it outside of the map and see if they're, you know, actually with this ESP, extra sensory perception hovering over things. There's all sorts of things you can do when you have control of the server as well, right? So, this idea that the server is an authoritative owner of the game state is pretty great when you own the game state.
So, you can put things in the game state to confuse the cheat. Um, and that's something that we like to do. And I I did this subtitle, too. LLMs aren't the only system that hallucinates. Lau, I'm sure this audience really appreciates that. They really appreciate your humor. It's my time to shine. Here's the architecture diagram, you guys. Here's the ready for this. Everybody take a picture of this. Post it on social media.
33:29 Anti-cheat data architecture
It's beautiful. I made this. If you take away one thing from this, it's that you could host the majority of your anti-gene processing on data bricks. You could if you wanted to and you should actually if I if I do say so myself. So if we work again, I don't know anymore. Uh if you look here, you'll actually see we have our sources uh for anti-che. You have your server side data coming in straight from your game server.
You have your client side data which is you know either your mouse or you know coming from a phone if you have a mobile game or coming from PCs. Uh and then we have some unstructured and semi-structured data as well. So like files and logs coming from say your kernel driver or crash logs from a server. Anything that's not fully structured you can use an ingest tool to handle your semi-structured or unstructured data.
And then you can actually just straight up stream live stream your uh client side and server side data. So if you can unify that, do it live streaming, then you can do things like live machine learning right on your uh live data stream which is pretty cool. So using ML flow or your data bricks notebooks, uh you can then serve your models in real time and get them out to make determinations. Um you can also let's see what else do we use here.
You can use delta live tables to get all of your uh ETL running and get everything in the way that it should be. Uh and then data brick SQL you can have your analysts actually use that as well to look back or look at matches once they finish you know usually running on about a 15inute uh roundtrip journey from a match to finish for the data to be available if you're not doing any sort of real-time streaming and then you hook that up to your BI tools.
You can make graphs like makes graphs and then you can get up here one day and show your graphs off trying not to fall asleep because they've been looking at graphs since what 11:00 a. m. this morning. We're sorry. We didn't want the last session. That was the one shilling slide. They don't call me Schillup Casinus for nothing. Shillup. Oh, I like that. I'm Schillup now. Uh, this is the fact that Pixelbots also exist on consoles.
Um, I won't spend a lot of time on this slide, but it's a vector. That's anything with a screen you can really do it on. And that is another one of your subtitles. I'm I know I feel like I'm not getting enough laughter from these. Like, guys, with a screen in a dream, you can cheat on anything. Oh, it's so good. Uh but yeah obviously uh I briefly visualized the two-dimensional input there and you can see that um the console's obviously uh some when the input is inhuman it kind of lights up the same way you can you can approach it the same this the same problem the same way DMA this is the next factor that we're seeing a lot of um this is basically uh a very niche way of hardware cheating it's it's growing in popularity it's been around for like 10 years but it's using physical hardware to read um read the physical memory of the of the computer uh with invisibly to the operating system.
35:53 DMA and hardware cheats
And um this is kind of what it looks like. You have you have your main PC that's running, you know, the anti-cheet and it's running Valerant. And then you leverage a DMA card over a PCIe DMA card to excfiltrate pages of uh of client memory to a second PC. It interprets those regions of memory. It tries to find the enemy positions and voila, you have a radar. Um and that's what baby's DMA looks like. Um wow, you're rushing me.
Uh we have three minutes left. We want people to ask questions. Okay. So, and then obviously there's ways to prevent this that often kind of uh we're talking about how much the whole reason it's viable is because the client has information that we you know it needs a little more information than a player should have. So even though it's not rendered them, it's still there. So uh trying to get a perfect fog of war system could completely eliminate the vector.
But like we were talking about before, we have to bake in some forgiveness into the game design or players would just pop in when they teleport if you don't actually know that they're coming. So to make sure it's really smooth and to make sure it feels really snappy to the end user, the client is given a little more information than it needs. You can also encrypt these objects, but that is honestly out of scope for this talk and is something that would probably uh confuse us.
Was that mine? All right, next. Uh so this is where they kind of go. It doesn't This is kind of where they go uh when they're total lunatics. Um could you imagine having more PCs than friends? Uh anyway, the they basically the same as before the PC's totally on you know the the entire cheat is implemented externally but you also leverage a KMB box just like in our pixelbot example and then because you have access to positions that you've excfiltrated from um from uh from memory directly.
You don't need to read the screen to find them. You have in world real positions. you need only translate those into 2D transformations for your mouse and then uh instrument it the same way you would a pixelbot and this I hate to say is something we see more and more every day um we really think the parents are responsible uh and DMA is growing in popularity it's something that uh we have developed we've spent a fortune on the capability to detect and there are ways to do it um but mostly what we do is actually something is a blocking pattern uh PCIe cards for the most part aren't necessary when you're playing a game.
38:20 Hardware mitigation and conclusion
I know that sounds strange. Um, but like if you think about a printer that's connected through a PCIe card or, you know, anything that isn't really a network adapter is probably not necessary. So, the majority of DMA firmwares can just be outright disabled. And we do that and that So, I'm not allowed to screenshot myself playing Valerant and then print it off immediately. No, I don't know why you would want to do that because I put screenshots on my fridge.
Ah, Cortana logged that one. That was a great kill. It's just like your dot matrix printer slowly fires up. But yeah, you can you can block a large percentage of these devices because a lot of them are very niche. Um and they really don't have any need to be operational while the game client's running. Siging sigging the device itself. Uh like it it obviously has um a config space that you can access with kernel antiche and you can try to identify something that is specific to a cheat.
But honestly these days the config space is honestly onetoone with a real DMA device. So it's very difficult to just outright detect a specific cheat that way. So uh most of the defense comes from like tightening the requirements on the host machine. So like uh leveraging things like TPM 2. 0 hypervisor code integrity. Um anyone here from Microsoft? Yeah. I mean we really you want to talk after this please. Uh but the problem problem with uh requiring these and MMU is another one. we can start to um leverage the the memory management unit to uh secure pages of memory from DMA reads.
But to do all this, we also need preboot DMA detection and kernel DMA protection to be enabled by default. And as a studio, we kind of inherit all that friction ourselves when we go to enforce it. And that's kind of like that risk loop I was showing out before. We try to selectively enforce these things in our most competitive and riskiest segments because it is a lot to ask somebody like could you enable secure boot especially like a a 15-year-old gamer.
It's like could you go into your bio settings don't mess up because that would be really bad and do all these things. Um and it's a lot of friction. So when it's not enabled by default us making that request is a lot. Uh and that's all right we have no time no time for questions. Um oh one more thing. This is worth saying. Give me one more. Okay I get one more. So this is the time to detection for any of those sheets we talked about during this whole presentation.
And as you can see, the ones that are external, the DMA and the pixelbot, they take us longer to detect because we're leveraging either nent, incredibly obscure hardware techniques to detect something that's a DMA card, or we're having to leverage a lot of aim sample data to know that you're using an external pixelbot. This makes them more po it lets them affect more matches because they're playing for longer. But it also means that in the infinite cat-and- mouse game that is anti-che, these types of cheats are slowly being um prioritized.
So, you know, you're in this talk because you're a data nerd. Now is your time to shine. Like, we did our job. Host antiche, we're done. I did it. It's secure. Now y'all got to detect mice. Uh or actually, you got to detect this. This is the last thing I'll leave you with.
Learn more about the Databricks Data and AI platform.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.