Efficiency Evaluation of Sparsified Homomorphic Encryption in Federated Learning

요약: 연합학습에 동형암호를 적용하면 프라이버시는 강해지지만 통신량과 연산량이 함께 늘어납니다. 이 포스터에서는 그 비용을 MNIST 기준으로 정량화하고, 희소화(Top-k)로 얼마나 회복할 수 있는지 측정했습니다. 평문 대비 동형암호는 통신량이 13.2배 늘었는데, 희소화를 적용하자 그 중 상당 부분이 되돌아왔습니다. 정확도 손실은 1.2%p 수준이었습니다.

⚡️ TL;DR

Homomorphic encryption makes federated learning private but expensive. We measured how expensive, and found that Top-k sparsification recovers most of the communication cost with a small accuracy loss.

👩🏻‍💻 My Role

• Designed and ran all three experimental conditions under a unified setup.
• Implemented the CKKS-based pipeline and the Top-k sparsification path.
• Analyzed results and wrote the paper.

🧪 Methods / Data

• Dataset: MNIST, evenly distributed across 10 clients.
• Conditions: (i) plaintext FL, (ii) CKKS-based HE-FL, (iii) HE-FL with Top-k sparsification (20%).
• Setup: up to 100 epochs, early stopping when accuracy improved less than 0.1%p for 10 consecutive epochs.
• Metrics: accuracy, training time, network traffic, and rounds to convergence.

📊 Results / Impact

• Encryption cost 2.8x training time and 13.2x communication over plaintext.
• Sparsification brought communication down 4.3x (2,198 → 512MB) and training time down 1.6x (214.1 → 133.0s).
• Accuracy dropped only 1.2%p against HE-FL, so the efficiency gain does not come at a meaningful cost in utility.
• This gap — the distance between a private system and a deployable one — is what my current work on sparsification-friendly CKKS aims to close.

Recommended citation: Jihyung Kook, Eunsang Lee (2025). "Efficiency Evaluation of Sparsified Homomorphic Encryption in Federated Learning." Proceedings of the KICS Fall Conference, pp. 752-753.
Download Paper