Biscotti: A Ledger for P2P Machine Learning Can federated learning be made more secure and private? We leverage blockchain primitives to enable secure peer-to-peer federated learning. A VRF committee verifies and aggregates SGD updates and appends the sum as a block in a ledger.
|
|
TorMentor: Private Multi-Party Machine Learning in an Untrusted Setting A distributed collaborative anonymous machine learning system that leverages Tor, with stronger privacy guarantees than modern solutions. This includes a side project that involves using gradient similarity to detect sybils in federated learning.
|
|
Distributed Clocks A vector clock instrumentation library. Distributed systems are difficult to analyze, and this project involves a suite of libraries for maintaining vector clocks and visualizing distributed communication. Includes support for Go, C, C++ and Java.
|
|
Samza StatServer @ Linkedin A system for collecting and aggregating user statistics for online machine learning, with a major emphasis on low read latency. Leverages Apache Kafka and Samza. Built on an internship at Linkedin in 2015.
|
|
Venice @ Linkedin Developed an early prototype of Linkedin's next generation derived data store. Leverages Apache Kafka and Voldemort DB. Includes a Hadoop to Venice data loading client. Built on an internship at Linkedin in 2014.
|