Efficient Privacy-Preserving Chatbot Learning through Prefix-Learning Based Federated Learning
Published in Conference on Information Security and Cryptography-Winter (CISC-W 2023), 2023
This paper proposes an efficient privacy-preserving approach for chatbot training using prefix-learning based federated learning. Our method significantly reduces communication overhead while maintaining strong privacy guarantees and model performance.
The proposed framework combines the advantages of federated learning for privacy preservation with prefix-learning techniques for efficient parameter updates. By training only a small subset of parameters (prefixes) locally and aggregating them federally, we achieve significant reductions in communication costs while preserving the privacy of sensitive conversational data. Comprehensive evaluations demonstrate the effectiveness of our approach across multiple chatbot benchmarks.
Recommended citation: Kyungju Choi, Byounghan Lee, Kyung-Ah Sohn. (2023). "Efficient Privacy-Preserving Chatbot Learning through Prefix-Learning Based Federated Learning." Conference on Information Security and Cryptography-Winter (CISC-W 2023), Seoul, South Korea.