Cheating humans already perform closely enough to trick such a system. Many cheaters are smart enough to use an aimbot only for a split-second to nail the flick. With a tiny bit of random offset, those inputs indistinguishable from a high-skill player.
These tricks may make it indistinguishable to a human moderator, but machine learning is actually really good at detecting that. But most companies don’t have the expertise, resources or training data to build a proper model for it.
Machine Learning is really good at CLAIMING it detected that.
The reality is that every few months there is a story about a fairly big streamer/e-sports player MAYBE getting caught cheating on stream. Sometimes it is obvious and sometimes it really becomes “Did they just know the map well enough to expect someone to come around that corner?”.
And a lot of times… it really is inconclusive. A somewhat common trope in movies is the veteran gunslinger literally aims at the wall of a stairwell and tracks where they expect the head to be and either fires a few rounds through the wall or waits for them at the bottom and… that is not entirely inconceivable considering that people tend to not crouch or move erratically down stairs. Obviously Jonathan Banks has a wallhack but Mike Ehrmantraut is just that damned good.
And false positives are a great way to basically kill a game. ESPECIALLY if they are associated with demonstrably false negatives too.
But you can be damned sure most of the major esports games are already doing this. It really isn’t expensive to train and they have direct feeds of every player in a tournament or twitch event. The issue is that there are (hopefully) tens of thousands of servers active at any moment and running Computer Vision+Inference on every single server is very costly.
And… I seem to recall there was a recent intentionally poorly defined Movement about maybe keeping user hostable dedicated servers a thing? How does that mesh with having every single server need to phone home (a fraction of) all 32 players feeds to a centralized cluster?
Machine learning doesn’t necessarily require a centralized cluster. Usually running those kinds of models is pretty cheap, it’s not an LLM basically. They usually do better than human moderators as well, able to pick up on very minute ‘tells’ these cheats have.
I understand your point about edge cases, but that’s not something the average player cares about much. E-sports is a pretty niche part of any game, especially the higher ranks. You just want to filter out the hackers shooting everyone each game that truly ruin the enjoyment. Someone cheating to rank gold instead of silver or whatever isn’t ruining game experiences; they’re usually detectable too, but if you get a false negative on that it’s not the end of the world. A smurf account of a very highly ranked player probably has a bigger impact on players’ enjoyment.
Depending on the model, inference can be run with CPU only. To distinguish what was originally proposed (a momentary flick consistent with aimbotting), you are either doing ray tracing (really expensive) or analyzing (effectively) video feeds. Both of which tend to put things more into the GPU realm which drastically increases the cost of a server.
But also? The only way these models can work is with constant data. Which means piping feeds back home for training which basically is never inexpensive.
Aside from that: if it was as simple as you are suggesting then this would be a solved problem. Similarly, if people don’t care about hackers outside of e-sports then there would be no reason for games to spend money on anti-cheat solutions when any match that matters would have heavy scrutiny. And yet, studios keep pumping out the cash for EAC and the like.
Cheating humans already perform closely enough to trick such a system. Many cheaters are smart enough to use an aimbot only for a split-second to nail the flick. With a tiny bit of random offset, those inputs indistinguishable from a high-skill player.
These tricks may make it indistinguishable to a human moderator, but machine learning is actually really good at detecting that. But most companies don’t have the expertise, resources or training data to build a proper model for it.
Machine Learning is really good at CLAIMING it detected that.
The reality is that every few months there is a story about a fairly big streamer/e-sports player MAYBE getting caught cheating on stream. Sometimes it is obvious and sometimes it really becomes “Did they just know the map well enough to expect someone to come around that corner?”.
And a lot of times… it really is inconclusive. A somewhat common trope in movies is the veteran gunslinger literally aims at the wall of a stairwell and tracks where they expect the head to be and either fires a few rounds through the wall or waits for them at the bottom and… that is not entirely inconceivable considering that people tend to not crouch or move erratically down stairs. Obviously Jonathan Banks has a wallhack but Mike Ehrmantraut is just that damned good.
And false positives are a great way to basically kill a game. ESPECIALLY if they are associated with demonstrably false negatives too.
But you can be damned sure most of the major esports games are already doing this. It really isn’t expensive to train and they have direct feeds of every player in a tournament or twitch event. The issue is that there are (hopefully) tens of thousands of servers active at any moment and running Computer Vision+Inference on every single server is very costly.
And… I seem to recall there was a recent intentionally poorly defined Movement about maybe keeping user hostable dedicated servers a thing? How does that mesh with having every single server need to phone home (a fraction of) all 32 players feeds to a centralized cluster?
Machine learning doesn’t necessarily require a centralized cluster. Usually running those kinds of models is pretty cheap, it’s not an LLM basically. They usually do better than human moderators as well, able to pick up on very minute ‘tells’ these cheats have.
I understand your point about edge cases, but that’s not something the average player cares about much. E-sports is a pretty niche part of any game, especially the higher ranks. You just want to filter out the hackers shooting everyone each game that truly ruin the enjoyment. Someone cheating to rank gold instead of silver or whatever isn’t ruining game experiences; they’re usually detectable too, but if you get a false negative on that it’s not the end of the world. A smurf account of a very highly ranked player probably has a bigger impact on players’ enjoyment.
Depending on the model, inference can be run with CPU only. To distinguish what was originally proposed (a momentary flick consistent with aimbotting), you are either doing ray tracing (really expensive) or analyzing (effectively) video feeds. Both of which tend to put things more into the GPU realm which drastically increases the cost of a server.
But also? The only way these models can work is with constant data. Which means piping feeds back home for training which basically is never inexpensive.
Aside from that: if it was as simple as you are suggesting then this would be a solved problem. Similarly, if people don’t care about hackers outside of e-sports then there would be no reason for games to spend money on anti-cheat solutions when any match that matters would have heavy scrutiny. And yet, studios keep pumping out the cash for EAC and the like.