Is there any work being done on trust-less (or maybe trust web) distributed model training? The main problem today is that training the model is being gatekept

Is there any work being done on trust-less (or maybe trust web) distributed model training? The main problem today is that training the model is being gatekept (is that a word?) by actors with 100's of k's of $. If there would be a way to run a client, like SETI@home, that will train models, then a few thousand unsophisticated users with 30x0's and some weeks/months of time will do to model training what bittorrent did to mp3 distribution. But for this to work, you need some way to feed images to users, ensure that images aren't re-used, somehow guard against malicious actors injecting faulty data etc.