[Paper Review] Privacy-preserving Federated Learning based on Multi-key Homomorphic Encryption

Jing Ma, Si‐Ahmed Naas+2 more|arXiv (Cornell University)|Apr 14, 2021

Privacy-Preserving Technologies in DataComputer Science51 references20 citations

TL;DR

This paper proposes xMK-CKKS, a privacy-preserving federated learning scheme using multi-key homomorphic encryption with an aggregated public key and collaborative decryption. It ensures model update confidentiality by requiring all participants to jointly decrypt aggregated results, resisting collusion attacks and reducing computational cost compared to Paillier-based schemes, while maintaining high model accuracy and low energy consumption (2.4W) suitable for IoT environments.

ABSTRACT

With the advance of machine learning and the internet of things (IoT), security and privacy have become key concerns in mobile services and networks. Transferring data to a central unit violates privacy as well as protection of sensitive data while increasing bandwidth demands.Federated learning mitigates this need to transfer local data by sharing model updates only. However, data leakage still remains an issue. In this paper, we propose xMK-CKKS, a multi-key homomorphic encryption protocol to design a novel privacy-preserving federated learning scheme. In this scheme, model updates are encrypted via an aggregated public key before sharing with a server for aggregation. For decryption, collaboration between all participating devices is required. This scheme prevents privacy leakage from publicly shared information in federated learning, and is robust to collusion between $k

Motivation & Objective

• To address privacy leakage in federated learning where model updates may expose sensitive data.
• To overcome the limitations of traditional homomorphic encryption in federated learning, where shared keys allow decryption by any participant.
• To design a multi-key homomorphic encryption scheme suitable for resource-constrained IoT devices.
• To ensure confidentiality of model updates even under collusion between k < N−1 devices and the server.
• To reduce computational and energy costs compared to existing homomorphic encryption-based federated learning schemes.

Proposed method

• Proposes xMK-CKKS, a modified version of MK-CKKS that uses an aggregated public key formed by summing individual public keys for encryption.
• Introduces a decryption share mechanism that combines individual secret keys and ciphertexts, ensuring no single participant can decrypt others’ data.
• Employs multi-key homomorphic encryption to allow homomorphic operations on encrypted model updates before aggregation at the server.
• Requires collaborative decryption involving all participating devices, preventing unauthorized access even if the server or some clients are compromised.
• Uses CKKS for approximate fixed-point arithmetic, enabling efficient computation on real-valued model weights.
• Deploys the scheme in a federated learning pipeline where devices encrypt gradients locally, send them to the server for aggregation, and jointly decrypt the global model update.

Experimental results

Research questions

• RQ1 Can a multi-key homomorphic encryption scheme be designed to prevent privacy leakage in federated learning when devices use different keys?
• RQ2 How can collaborative decryption be achieved without requiring interaction among devices during encryption or aggregation?
• RQ3 Can the proposed scheme reduce computational and energy costs compared to Paillier and MK-CKKS-based federated learning?
• RQ4 Does the scheme maintain model accuracy comparable to non-encrypted federated learning in real-world IoT scenarios?
• RQ5 How robust is the scheme against collusion between a subset of devices and the server?

Key findings

• The xMK-CKKS-based federated learning scheme preserves model accuracy comparable to traditional federated learning and secure schemes using MK-CKKS or Paillier encryption.
• Computational cost is significantly lower than Paillier-based federated learning, with performance close to MK-CKKS across all phases (encryption, aggregation, decryption, decryption share generation).
• Energy consumption averages 2.4 Watts across 10 Jetson Nano devices, demonstrating suitability for low-power IoT deployments.
• The scheme is robust against collusion between k < N−1 devices and the server, as decryption requires all participants’ shares.
• Communication cost remains low, as only encrypted model updates are transmitted, not raw data.
• The scheme enables secure, scalable, and efficient federated learning in smart healthcare scenarios using edge devices.