Advancing Federated Learning through Verifiable Computations and Homomorphic Encryption.

Area of Science:

• Privacy-preserving machine learning
• Applied cryptography
• Distributed systems

Background:

• Federated learning (FL) is vulnerable to malicious nodes and global model privacy breaches.
• Existing FL research lacks methods to protect the global model's privacy.
• Emerging cryptographic tools like ZKVM and homomorphic encryption offer potential solutions.

Purpose of the Study:

• To introduce a novel federated learning framework enhancing security and privacy.
• To address the challenge of protecting global model privacy in FL.
• To leverage ZKVM for verifiable proofs and homomorphic encryption for secure computation.

Main Methods:

• Integration of Zero-Knowledge Virtual Machine (ZKVM) for local computing provers and execution integrity proofs.
• Implementation of Fully Homomorphic Encryption (FHE) to enable computation in the ciphertext space.
• Development of a new federated learning framework combining ZKVM and FHE.

Main Results:

• Achieved zero-knowledge proofs (ZKP) for multi-class and arbitrarily scaled machine learning tasks.
• Ensured global model privacy during local computation and transmission via encryption.
• Demonstrated a satisfactory 90% model accuracy on the IRIS dataset despite FHE's impact.

Conclusions:

• The proposed framework offers a highly secure approach to federated learning.
• ZKVM and FHE integration effectively protects global model privacy and ensures computational integrity.
• The framework shows promise for future improvements in efficiency and security as cryptographic tools advance.