Leverage Variational Graph Representation for Model Poisoning on Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Federated Learning

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data.
• Model poisoning (MP) attacks threaten FL integrity by manipulating local models.
• Existing attacks often require access to training data or are detectable.

Purpose of the Study:

• Introduce a novel training data-untethered model poisoning attack on federated learning.
• Develop a method that is both effective and elusive to detection.
• Assess the threat posed by this new attack to FL systems.

Main Methods:

• Extended an adversarial variational graph autoencoder (VGAE) for model poisoning.
• Developed the VGAE-MP attack, leveraging overheard benign local models without training data access.
• Implemented a new attacking algorithm using VGAE and sub-gradient descent for optimal benign model selection.

Main Results:

• The VGAE-MP attack effectively degrades federated learning accuracy.
• Existing defense mechanisms proved ineffective in detecting the VGAE-MP attack.
• The attack demonstrates a severe threat to the security and reliability of FL.

Conclusions:

• The proposed VGAE-MP attack presents a significant advancement in federated learning security threats.
• The data-untethered nature and elusiveness of VGAE-MP highlight vulnerabilities in current FL defenses.
• Further research into robust defense strategies against such sophisticated attacks is crucial.