SIA: A sustainable inference attack framework in split learning.

Area of Science:

• Computer Science
• Machine Learning
• Cybersecurity

Background:

• Split learning is a distributed framework for resource-constrained joint training.
• Existing feature space hijacking attacks pose privacy risks by reconstructing client data.
• Current defenses may not adequately protect against sophisticated attacks.

Purpose of the Study:

• To develop an enhanced attack framework for efficient data reconstruction in split learning.
• To extend data reconstruction attacks to vertically partitioned data scenarios.
• To create an attack that evades state-of-the-art detection mechanisms.

Main Methods:

• Developed an enhanced feature space hijacking attack.
• Adapted the attack for vertically partitioned data in split learning.
• Introduced three attack training modes for flexibility.
• Evaluated attack effectiveness, invisibility, and generality across datasets and defenses.

Main Results:

• The enhanced attack achieves high-quality data reconstruction with minimal impact on the main task.
• The attack successfully evades current state-of-the-art detection mechanisms.
• The attack is effective and generalizable across various datasets and scenarios.
• The attack is successfully extended to the challenging vertically partitioned data setting.

Conclusions:

• Split learning faces significant, underestimated privacy risks from advanced reconstruction attacks.
• The developed attack framework demonstrates effectiveness, stealth, and adaptability.
• Urgent need for robust security measures in split learning applications is highlighted.