Goalie: Defending Against Correlated Value and Sign Encoding Attacks.

Area of Science:

• Machine Learning Security
• Data Trust Security
• Privacy-Preserving Machine Learning

Background:

• Data trusts face correlated value and sign encoding attacks to steal shared data.
• Current defenses degrade benign model performance by perturbing parameters or data.
• A need exists for methods that protect data without compromising model utility.

Purpose of the Study:

• To propose Goalie, a novel method to defend data trusts against data theft attacks.
• To detect and stop malicious model training while preserving benign model performance.
• To ensure the integrity and security of shared data within data trusts.

Main Methods:

• Goalie detects malicious models by analyzing changes in parameter distributions caused by encoding attacks.
• Features are extracted from model parameters during early training epochs.
• Theoretical analysis confirms that regularization terms used in attacks alter parameter distributions.

Main Results:

• Goalie achieves high accuracy (over 0.9) in detecting models with encoding attacks.
• The method demonstrates effectiveness even in extreme attack scenarios.
• Detection is highly efficient, taking only 1.1 ms using early-epoch features.

Conclusions:

• Goalie effectively defends data trusts against sophisticated data theft attacks.
• The proposed detection mechanism preserves benign model performance.
• Goalie offers a computationally efficient and accurate solution for data trust security.