Survey: Leakage and Privacy at Inference Time

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Publicly available machine learning (ML) models pose significant data leakage risks.
• Commercial and government ML applications may involve sensitive user and client data.

Purpose of the Study:

• To provide a comprehensive survey of contemporary advances in ML data leakage.
• To cover involuntary, malicious, and defense mechanisms against data leakage.
• To focus on inference-time leakage in publicly available models.

Main Methods:

• Discussed the nature of data leakage in various contexts (data, tasks, model architectures).
• Proposed a taxonomy for involuntary and malicious leakage.
• Described current defense mechanisms, assessment metrics, and applications.

Main Results:

• Categorized data leakage into involuntary and malicious types.
• Reviewed existing defense strategies and evaluation metrics.
• Identified key challenges and future research avenues.

Conclusions:

• Data leakage from ML models is a critical concern requiring robust defenses.
• Further research is needed to address outstanding challenges in ML model security.
• Understanding leakage types and defenses is crucial for secure ML deployment.