III / V2024SpartaHack 9

FedUp

Many hands, no shared data.

Peer-to-peer federated learning: nodes pass a model around the network, train on data they never share, and converge in fewer epochs than a traditional model. Later the seed of an arXiv paper.

BEST HACK · ALL TRACKSRead the room

1stoverall, 350+ hackers
0 bytesof training data shared

Inspiration

Over half of enterprise data goes unused because of legal and security hurdles. We wanted a way to aggregate insight from data while the data itself stays private.

What it does

Clients join a peer-to-peer network and train one shared model without ever sharing training or test data. Only gradient updates move. On three team members’ devices it beat a traditional model at predicting loan defaults, in epochs to convergence and in validation loss.

How we built it

Nodes register with a discovery server, which broadcasts a topology and picks a starting node. From there training is fully decentralized: each node takes the model from its predecessor, trains locally, and passes it on. PyTorch over TCP/IP.

III / V2024SpartaHack 9

FedUp

Many hands, no shared data.

Peer-to-peer federated learning: nodes pass a model around the network, train on data they never share, and converge in fewer epochs than a traditional model. Later the seed of an arXiv paper.

BEST HACK · ALL TRACKSRead the room

1stoverall, 350+ hackers
0 bytesof training data shared

Inspiration

Over half of enterprise data goes unused because of legal and security hurdles. We wanted a way to aggregate insight from data while the data itself stays private.