MaskCRYPT: Federated Learning With Selective Homomorphic Encryption

Summary:
A review of MaskCRYPT, which proposes selective homomorphic encryption to balance privacy and efficiency in federated learning.

🔑 Research Question:

• Do we need to encrypt all model weights to ensure privacy?

⚙️ Key Mechanism: Selective Homomorphic Encryption

• Uses gradient-based importance to identify which parameters should be encrypted
• Each client generates a local mask based on importance scores
• The server aggregates these into a global Mask Consensus
• Only selected parameters are encrypted, while others are aggregated in plaintext

📊 Main Results:

• Encrypting as little as 1% of parameters can defend against membership inference and reconstruction attacks
• Reduces communication overhead by up to 4.15×
• Improves wall-clock training time
• Maintains accuracy comparable to non-encrypted training

⚠️ Limitations / Open Questions:

• Requires communication of local importance scores
• Fairness and correctness of Mask Consensus must be ensured
• Evaluated only on moderate-scale models and datasets

💡 Insight:
MaskCRYPT suggests that full encryption may be unnecessary in practical federated learning systems, and that selectively protecting sensitive components can provide a better trade-off between security and efficiency.

Slides:
PDF (Korean) Download