search

waldo-vision / optical.flow.demo

A project that uses optical flow and machine learning to detect aimhacking in video clips.
https://www.waldo.vision
Mozilla Public License 2.0
536 stars 24 forks source link

Doesn't success of this project screw itself over? #26

Closed entropylost closed 2 years ago

entropylost commented 2 years ago

Let’s assume this project succeeds in detecting which patterns of movement are “human” and which ones aren’t. What’s to say that the hackers can’t just use this program to make their bots follow exactly what’s “human”. (I don’t know anything about machine learning so I have no idea whether this is trivial)

hippiewho commented 2 years ago

Well I would imagine the training data could be controlled. The analysis functionality probably shouldnt be used to train the model.

jaredb1011 commented 2 years ago

Partially, sure. But it forces cheat devs to stay on their game and keep working. There's always going to be new cheats, and new anti-cheats.

dorp92 commented 2 years ago

Using generative adversarial networks like architecture will screw this project easily. This anti-cheat is basically a discriminator that determine if hacks used or not. The hackers will need to train a "generator" (hack) that fool the discriminator (anti cheat). Good GANs have been show to produce realistic images that even humans can't defer from real once. This is a core issue and I don't think that more development work will solve it.

v0idp commented 2 years ago

there is also the possibility of people actually hacking the NN itself by messing with the video stream. this was already proven to work with object recognition by creating custom noise and adding it to the frame. sometimes even one pixel only. but the future will tell honestly.

see: https://www.youtube.com/watch?v=xHpwLiTieu4