Initial Project Steps

[jaredb1011]

• Design a program that can cleanly extrapolate mouse movement data from gameplay footage.
• Begin to train deep learning models on the differences between human gameplay, and machine assisted gameplay.
• Research existing programs/libraries that could be useful or simplify our program.

[vkottler]

Is "an AI can be trained to categorize video data as having 'human' or 'software' input-control" really a valid premise?

If software created to cheat in games is already sophisticated enough to bypass game-developer anti-cheat mechanisms, wouldn't it mostly trivial to add characteristics like:

• Slew-rate limits on movement
• Variable-frequency jitter
• Periodic intentional error
• Randomized reaction delay

And other obfuscation features that will make it unlikely for trained AI to tell the difference?

Is it really even possible to glean game-model data from a video stream to a level of fidelity that it can be plumbed to AI and still be useful?

I think what you're trying to do here is probably not possible, and the point of saying that is not to be discouraging - I just wonder if there's another way to go about this.

It seems a bit irresponsible to advertise "a project" to solve this kind of problem that's at the maturity level of "two Python files" ...

[Monstrous-Pro]

These are all valid points and things I also thought of at first, but the beauty of machine learning is that it can still have reasonable error if given enough quality data. Emphasis on quality, since a lot of things can get messy with strictly image data if there’s not high enough resolution, and if there’s not enough properly labeled data, i.e. confirmed cheating/not cheating gameplay.

It’s still very early but ML shouldn’t be underestimated.