Related Experiment Videos
SLeak: Multi-Target Privacy Stealing Attack Against Split Learning
IEEE Transactions on Pattern Analysis and Machine Intelligence
|January 14, 2026
Summary
Split Learning (SL) is vulnerable to privacy attacks. A new threat, Split Leakage (SLeak), exploits client representation preferences to steal data and functionality, even with limited public data.
Area of Science:
- Artificial Intelligence
- Machine Learning Security
- Distributed Systems
Background:
- Split Learning (SL) offers privacy and efficiency but faces inference attack risks.
- Existing SL privacy defenses rely on unrealistic assumptions, limiting real-world effectiveness.
- Server adversaries can potentially compromise client privacy in SL frameworks.
Purpose of the Study:
- To investigate inherent vulnerabilities in Split Learning (SL) frameworks.
- To introduce a novel privacy threat, Split Leakage (SLeak), against SL.
- To demonstrate SLeak's effectiveness without strong privacy assumptions.
Main Methods:
- Analyzing client representation preferences in SL's smashed data and server model.
- Developing a substitute client to mimic target client behavior.
- Introducing the Split Leakage (SLeak) threat for multiple privacy objectives.
- Utilizing partial same-domain auxiliary public data for attacks.
Main Results:
- Identified that both smashed data and server models reveal client representation preferences.
- Demonstrated that a substitute client can perfectly replicate target client functionality, data, and labels.
- SLeak attacks showed superior performance compared to state-of-the-art methods across diverse datasets and models.
- Ablation studies confirmed SLeak's robustness and applicability in various scenarios.
Conclusions:
- Split Learning (SL) possesses inherent vulnerabilities exploitable by server adversaries.
- Split Leakage (SLeak) presents a practical and effective threat to SL privacy.
- SLeak's success with minimal data requirements highlights significant privacy risks in distributed learning.
