Notes form the 7/14/2022 meeting of TFF collaborators

  • Participants: Krzysztof Ostrowski (Google), Boyi Chen (LinkedIn)

  • Boyi’s update on LinkedIn’s progress and plans.

    • Onboarded to TFF and integrated into ML infra
    • Doing offline experiments on use of TFF for enterprise solutions
    • Three areas of interest
      • Freerider attacks
        • Someone wants to contribute zeros, reap benefits
        • Two goals - detection, solutions
        • Model poisoning a distinct goal, but seemingly related
      • Bias with heavily skewed contributors
        • Some contributors having much more data than others
        • Goes both ways - heavy users over-influencing the model, but also lots of lightweight users dragging performance down
      • Cross-silo FL for a mixture of data from LinkedIn and from outside
        • Guarantees on data not mixing
      • Simulations of on-device FL
        • Simulation capability already exists - we’re talking about simulating the behaviors seen in a realistic prodution environment
        • Vary distributions of things like device processing power to asses how it may impact training performance
    • Currently not much progress running on Azure, so punt on this for now
  • Modes of contributing / working together:

    • Algorithms and coimponents in TFF for detecting freeriders and mitigating that
      • Design doc - loop in people from both ends to help improve
      • LinkedIn could contribute code
      • Tentatively LinkedIn to own or co-own a directory within TFF repo where this could go - tbd whether one or more of these and where they would go
  • TFF’s plans

    • Empower partners to build platforms based on TFF
      • Components
      • References architectures
      • Both cross-silo and cross-device
        • Some code is already in OSS, more code upcoming
      • End-to-end privacy, etc., guarantees for platform partners
  • Next steps:

    • Create individual proposals to iterate on with people from both sides
    • Prioritize together
      • Maybe that means increasing frequence to once per 2 weeks
      • Pick topics to unpack, loop in people interested in the topic