Bigger and better actually worked pretty well w.r.t. MAD. Now we are developing smaller and better, which seems almost more dangerous since we're already using drones more flippantly than conventional military arms.

Maybe to handle knockback you could allow recoiling rotation around one axis and have the rotors automatically anticipate the force on fire.

Why do you need machine learning to "deal with the knockback of shooting a gun from a drone"?

It is very easy to work on these tools without even realizing it. Let us say that you are developing a video processing algorithm and publish the code on github. The developers of missile software can and will use it. At some point, they may ask you for consulting/support in exchange for a large amount of money and you will say yes, and never realize who they were. I'm not saying that this has happened several times to my lab.

In the video the drone self-destructs by exploding and producing shrapnel, I doubt knockback is a problem.