Back to the Drawing Board - A Critical Evaluation of Poisoning Attacks on Federated Learning

Explore a critical evaluation of poisoning attacks on federated learning in this 20-minute IEEE conference talk. Delve into traditional machine learning, cross-device FL, and various poisoning attack strategies. Examine key questions, prior work, and three main dimensions of attacks. Analyze global model parameters, model poisoning, and practical threat models. Gain insights into untargeted attacks, data poisoning, and key results across different federated learning scenarios. Evaluate the robustness of federated learning systems and understand the implications for both cross-silo and cross-device implementations.

Introduction
Traditional Machine Learning
CrossDevice FL
Poisoning Attacks
Literature
Key Question
Outline
Prior Work
Three Main Dimensions
Global Model Parameters
Model Poisoning
Takeaways
Impractical Threat Models
Most Severe Threat Model
Untargeted Attacks
Practical Threat Models
Intuition
Data Poisoning
Key Results
Nonrobust Federated Learning
Cross Silo Federated Learning
CrossDevice Federated Learning
Robustness of Federated Learning