Outline
Tentative Outline
Part I: What is Federated Learning and Analytics
Overview
Characteristics
Cross-device vs. cross-silo FL
FL vs. datacenter distributed ML
FL vs. fully decentralized Learning
Federated analytics
Part II: Federated Learning, Optimization, and Efficiency
Federated Averaging (FedAvg)
Problem formulation
Challenges in FL
Optimization methods for FL
Communication-efficient FL
FedAvg in TensorFlow Federated (TFF)
Good baseline tasks and benchmarks
Best practices and newer algorithms
Hyperparameter tuning challenges
Part III: Privacy for Federated Learning and Analytics
Overview
Actors, threat models, and privacy in depth
Protections against external compromised actors
Protections against a compromised server
Differentially Private Federated Training
Central vs. local vs. distributed differential Privacy
Example- vs. user-level differential privacy
Differentially Private FedAvg
Part IV: Open Problems
Modeling heterogeneous data
Personalization
Connections to meta-learning, multi-task, transfer learning
Fairness and bias
Robustness
Goals and capabilities of an adversary
Types of adversarial attacks
Defense mechanisms and open challenges
System challenges
Cross-silo (e.g. heterogeneous data schemas)
cross-device (availability, reliability)