Now you're trading a prediction problem for an implementation problem as I explained in another comment. A few heavily populated counties is not a small attack

Now you're trading a prediction problem for an implementation problem as I explained in another comment. A few heavily populated counties is not a small attack surface. You need to hack hundreds of airgapped optical scanners in each state (and they all use different machines), not too much or too little, because then you'll trigger a recount using the real paper ballots. You still need basically perfect prediction to win and stay undetected, and you're up against big election departments with tons of employees and lots of integrity checks.