Privacy and Robustness in Federated Learning: Attacks and Defenses

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Centralized AI model training faces challenges due to data silos and privacy concerns.
• Federated learning (FL) emerges as a decentralized alternative, but existing protocols are vulnerable to attacks.
• Ensuring data privacy and system robustness is crucial for the widespread adoption of FL.

Purpose of the Study:

• To provide a comprehensive survey of privacy and robustness in federated learning over the past five years.
• To introduce a unique taxonomy of FL threats, including privacy attacks and poisoning attacks.
• To highlight key techniques, assumptions, and future research directions in robust and privacy-preserving FL.

Main Methods:

• Literature review and synthesis of research on federated learning privacy and robustness.
• Development of a taxonomy categorizing threat models, privacy attacks/defenses, and poisoning attacks/defenses.
• Analysis of intuitions, techniques, and assumptions underlying various FL security measures.

Main Results:

• Existing FL protocols are vulnerable to internal and external adversaries, compromising data privacy and system integrity.
• A structured overview of privacy attacks (e.g., inference, membership) and defenses (e.g., differential privacy, secure aggregation) is presented.
• Poisoning attacks (e.g., data, model poisoning) and their corresponding defenses are analyzed, revealing critical vulnerabilities.

Conclusions:

• There is a critical need for robust and privacy-preserving FL systems to mitigate diverse adversarial threats.
• Future research should focus on developing advanced defense mechanisms and understanding the interplay between privacy, robustness, and FL's multidisciplinary goals.
• Continued investigation into threat models and attack/defense strategies is essential for advancing secure federated learning.