Analyzing the vulnerabilities in Split Federated Learning: assessing the robustness against data poisoning attacks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Centralized machine learning faces privacy challenges.
• Distributed Collaborative Machine Learning (DCML) offers privacy-preserving alternatives.
• Split Federated Learning (SFL) integrates Split Learning (SL) and Federated Learning (FL).

Purpose of the Study:

• To comprehensively study and analyze the impact of data poisoning attacks on Split Federated Learning (SFL).
• To propose and evaluate novel attack strategies against SFL systems.

Main Methods:

• Developed three data poisoning attack strategies: untargeted, targeted, and distance-based.
• Evaluated attacks on Electrocardiogram Signal Classification and MNIST dataset.
• Varied malicious client percentages and model split layers.

Main Results:

• Distance-based and untargeted attacks demonstrated a greater impact on classifier performance degradation.
• Attack effectiveness was analyzed concerning varying malicious client ratios and split points.
• Targeted attacks showed less impact compared to distance-based and untargeted strategies.

Conclusions:

• Split Federated Learning is vulnerable to various data poisoning attacks.
• Distance-based and untargeted attacks pose a significant threat to SFL classifier integrity.
• Understanding these vulnerabilities is crucial for developing robust SFL defenses.