(2019-04-15, 05:52 PM)JonasP Wrote: How's the performance hit on the server?
The networks and the models are quite small and take few tens of microseconds per detection on my computer. I had anticipated complications (which didn't happen) and prematurely adding threading support. Most of the computation work happens in different threads which you can configure. Hence, the SAMP server is never blocked by the plugin.
In a nutshell, the performance hit is near zero. You can do 25,000 detections every second.
(2019-04-15, 05:52 PM)JonasP Wrote: Any plans to expand into other weapons?
Yes, but I don't want to add a zillion weapons and let all fail. The weapons with high fire rate make it relatively easier to detect aimbots. Hence, I started off with those.
(2019-04-15, 05:52 PM)JonasP Wrote: It would be interesting to try this out and give the data to admins so they could monitor various situations but most firefights here happen with shotguns, deagles, snipers, combat shotguns, fire extinguishers. C-bug is also used pretty much always. So I'm wondering if there is a possibility that this plugin will receive updates and include these weapons take cbug and cslide into account.
After hundreds of man-hours of research, I am reasonably certain that it's difficult to detect aimbots when used with slow firing weapons such as desert. I need to experiment much more before making a detector for deagle or shotgun.
I have made some progress with sniper but haven't published it yet.
The focus currently is on to get the detectors to work really well on M4/AK/MP5.
(2019-04-15, 05:52 PM)JonasP Wrote: Does it build some sort of model around a specific playerid or for all players and then looks at the odd results?
If a player disconnects and reconnects to a new playerid, does the AI need to spend time relearning about that player? Because if it builds data for specific playerids maybe it would be possible to collect that data and save it so that when the player joins the next day it just loads that data and continues learning about him instead of having to learn everything from zero again?
There is a common model for all the weapons and for all the player. The models were trained using dataset from around ten people with varying skills. Three skin aimbots were used for training. Hence, I speculate that there will be situations where super skilled players or weirdos make the detectors throw out false positives. It's speculation. It never happened when it was tested by several people but this should be taken with a grain of salt because the same people who were testing were also the ones who gave the training data.
I have been waiting for people to use and report issues which no one has done yet. I can't do much without getting feedback. I don't own a server and cannot test it.