EDLaaS:Fully Homomorphic Encryption over Neural Network Graphs for Vision and Private Strawberry Yield Forecasting

Area of Science:

• Computer Science
• Cryptography
• Machine Learning

Background:

• Privacy-preserving machine learning (PPML) is crucial for sensitive data.
• Fully Homomorphic Encryption (FHE) allows computation on encrypted data, but its application in deep learning is complex.
• Existing FHE schemes often lack user-friendliness for deep learning integration.

Purpose of the Study:

• To present an automatically parameterized FHE framework for encrypted neural network inference.
• To enhance the usability and applicability of FHE in deep learning contexts.
• To demonstrate the potential of encrypted deep learning in real-world sensitive applications.

Main Methods:

• Utilized the fourth-generation Cheon, Kim, Kim, and Song (CKKS) FHE scheme over fixed points.
• Employed the Microsoft Simple Encrypted Arithmetic Library (MS-SEAL).
• Developed an open-source framework with reproducible examples for FHE-compatible neural network inference.

Main Results:

• Achieved enhanced usability and applicability of FHE for neural network inference.
• Demonstrated that FHE is suitable for private predictions but has limitations (e.g., model training).
• Showcased competitive performance in strawberry yield forecasting using encrypted deep learning.

Conclusions:

• FHE can enable completely private neural network predictions in specific contexts.
• Lowering barriers to entry for FHE can benefit sensitive fields by allowing use of third-party neural networks.
• Encrypted deep learning adoption can increase AI use in privacy-conscious sectors like agri-food, supporting net-zero goals.