/srs_talk_aug4

Summer Research Seminar talk on August 4

FedCampus: A Federated Learning & Federated Analytics Experiment on DKU Campus

Motivation for FedCampus (Steven, 20s)

How do we learn from data generated on DKU campus, without collecting them?

(Ask audience for answers).

(This is exact what FedCampus aims to achieve).

Introducing FedCampus (30s)

  • Privacy-preserving data collection platform for smart campus.
    • Data analytics.
    • Benefits community.
  • Edge devices: smartphones, smartwatches, IoTs.
  • Systems development to provide research potentials.
    • Some of them we introduce next.
  • Collaboration wanted.

(Remind the question).

Steps to build up FedCampus

  1. Federated Learning platform.
  2. ML & FL algorithms.
  3. User-facing app.

Federated Learning platform

Introducing Federated Learning (FL) (1min)

  • Local data & local ML training.
  • Central server coordinate training and aggregate ML model.
  • Analogy: the Federated Government.

FL example: Gboard in 2017 (20s)

Google's smart keyboard.

  • Next word prediction.
  • Train on user's phones when idle.

Why FL is the way forward (20s)

  • Centralized ML invades privacy land.
  • ML using privacy data is useful.
  • FL solves this problem, especially mobile FL.

Existing mobile FL solutions suck (20s)

  • FLaaS: send data to proprietary cloud.
  • Open source solutions: poor mobile support/ very basic.

FedKit: on-demand mobile FL platform for FedCampus

  • Self-hosted & open source.
  • Persistent on-demand service.
  • Server-side ML model swapping.
  • Telemetry.

(Tech stack graph).

FedKit training process (1min)

  1. ML model: app obtains .tflite model from backend.
  2. Spawn Flower server: app requests backend.
  3. Train: using Flower with gRPC connection.

ML & FL algorithms

Health data and privacy

  • Tremendously useful yet sensitive use case.

Case study: sleep efficiency prediction (Aicha)

Dealing with non-IID-ness (Tianjun)

TODO: How does this fit in.

User-facing app (Beilong)

TODO: Getting data from Huawei Health, etc.

Physical infrastructure (Johnny)

TODO: We cannot have data safety using cloud.

How you can engage with us

  • We are looking for collaboration using FL.
  • We are continuously looking for new members.
  • We will soon be looking for participants in our experiments.