Privacy for AI from NP-Hard Problems - Universal Compute on Encrypted Data

Overview

Explore cryptographic methods for private AI, including MultiParty Computation, Threshold Cryptography, and Fully Homomorphic Encryption, with focus on challenges in evaluating operations like softmax in the encrypted domain.

Syllabus

• Introduction to Privacy in AI
Overview of Privacy Concerns in AI
Importance of Cryptographic Techniques
• Cryptographic Foundations
Overview of Cryptography
Basic Cryptographic Primitives
• MultiParty Computation (MPC)
Definition and Principles of MPC
Common Protocols and Applications
Real-world Use Cases and Challenges
• Threshold Cryptography
Introduction to Threshold Schemes
Designing Threshold Cryptographic Protocols
Security and Performance Considerations
• Fully Homomorphic Encryption (FHE)
Principles of FHE
Current FHE Schemes and Their Implementations
Evaluating Efficiency and Use Cases
• Privacy-preserving AI Operations
Challenges in Encrypted Domain Computation
Specific Challenges in Encrypted Softmax Evaluation
Techniques for Neural Network Evaluation on Encrypted Data
• Designing Privacy-preserving AI Systems
Integrating Cryptographic Methods into AI Workflows
Overcoming Practical Implementation Barriers
Case Studies and Real-world Examples
• Advanced Topics and Emerging Trends
Hybrid Approaches in Privacy-preserving AI
Emerging Cryptographic Techniques for AI
• Conclusion and Future Directions
Summarizing Key Learnings
Potential Future Developments in AI Privacy
• Final Project or Assessment
Design a Privacy-preserving AI Prototype
Evaluate and Present a Case Study on Encrypted AI Computations

Subjects

Computer Science