Privacy-Preserving Decision-Tree Evaluation with Low Complexity for Communication

Area of Science:

• Cryptography
• Machine Learning Security
• Privacy-Preserving Computation

Background:

• Companies offer machine learning services, raising concerns about user data and model privacy.
• Existing privacy solutions are communication-intensive and vulnerable to quantum attacks.

Purpose of the Study:

• To design a secure integer-comparison protocol and a client-server classification protocol for decision-tree evaluation.
• To develop a privacy-preserving, quantum-resistant, and communication-efficient machine learning classification method.

Main Methods:

• Developed a secure integer-comparison protocol utilizing fully homomorphic encryption.
• Proposed a client-server decision-tree classification protocol based on the secure integer-comparison protocol.
• Implemented a lattice-based fully homomorphic encryption scheme resistant to quantum attacks.

Main Results:

• The proposed classification protocol achieves a low communication cost, requiring only one round of user interaction.
• Experimental results demonstrate a communication cost reduction to 20% compared to traditional methods.
• The protocol is built on a quantum-resistant lattice-based fully homomorphic encryption scheme.

Conclusions:

• The novel protocol offers a secure, efficient, and quantum-resistant solution for privacy-preserving machine learning classification.
• This approach significantly lowers communication overhead in client-server machine learning interactions.
• The use of lattice-based fully homomorphic encryption provides a robust defense against future quantum computing threats.