Respecting Partial Privacy of Unstructured Data via Spectrum-Based Encoder.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine Learning as a Service (MLaaS) is increasingly popular.
• Users risk exposing sensitive, non-task-related information in uploaded data.
• Existing methods for filtering task-independent information are challenging for unstructured data.

Purpose of the Study:

• To propose a novel framework for protecting user privacy in MLaaS.
• To effectively filter task-independent information from unstructured data.
• To maintain the utility of task-related information for inference.

Main Methods:

• A spectrum-based encoder transforms unstructured data into a latent space.
• A task-specific model identifies essential information for the target task.
• The framework was evaluated on three benchmark visual datasets.

Main Results:

• The proposed framework offers superior protection for task-independent information.
• The system effectively maintains the usefulness of task-related information.
• Performance was compared against previous works on benchmark datasets.

Conclusions:

• The novel framework successfully addresses privacy leakage risks in MLaaS for unstructured data.
• This approach enhances user privacy without compromising the effectiveness of machine learning tasks.
• The method provides a robust solution for privacy-preserving MLaaS.