disco
disco copied to clipboard
epfml
Decentralized & federated privacy-preserving ML training, using p2p networking, in JS
This is a dummy PR to test the workflow. I'll update it once changes on Federated Learning to implement WS are completed.
While the DEV.md gives some hints as how to get started and where some things can be found, it is not very beginner friendly; a new student would struggle with...
It would be great to add [tensorboard](https://js.tensorflow.org/api_node/1.0.2/) to more easily plot the output of the benchmark CLI. The basic idea is to use [tf.node.summaryFileWriter](https://js.tensorflow.org/api_node/1.0.2/#node.summaryFileWriter) to log the output in the...
Currently, a user who wishes to create a new task doesn't find any information to the parameters that they must / may set. There are several ways to fix this,...
In decentralized learning with secure aggregation, the helper server can keep track of who aggregates with whom. Thus, the server can easily collect statistics like the number of aggregation rounds...
We should explicitly set values for `scheme, noiseScale, clippingRadius, decentralizedSecure, minimumReadyPeers`, and `maxShareValue` in all existing tasks ([here](https://github.com/epfml/disco/tree/develop/discojs/src/tasks)). Let's turn secure aggregation on and set `clippingRadius` to `20` and `maxShareValue`...
**Describe the bug** The decentralized router has a signaling server that is used for managing all communication. The same instance of signalingServer is used for all decentralized tasks that are...
We would like to implement secure aggregation MPC in federated learning (similar to issue (https://github.com/epfml/disco/issues/9) for decentralized learning). Through secure aggregation, the server will no longer be able to reconstruct...
