Towards Confidential Chatbot Conversations: A Decentralised Federated Learning Framework

Abstract

The development of cutting-edge large language models such as ChatGPT has sparked global interest in the transformative potential of chatbots to automate language tasks. However, alongside the remarkable advancements in natural language processing, concerns about user privacy and data security have become prominent challenges that need immediate attention. In response to these critical concerns, this article presents a novel approach that addresses the privacy and security issues in chatbot applications. We propose a secure and privacy-preserving framework for chatbot systems by leveraging the power of decentralised federated learning (DFL) and secure multi-party computation (SMPC). Our DFL framework leverages blockchain smart contracts for participant selection, orchestrating the training process on user data while keeping the data local, and model distribution. After each round of local training by the participants, the blockchain network securely aggregates the model updates using SMPC, ensuring that participants’ raw model parameters are not exposed to others. The global model is encrypted and stored in hypermedia protocols such as the InterPlanetary File System. Participants decrypt the global model updates using their private keys to further fine-tune their models. Iterative training rounds are executed through the blockchain network, with participants updating the model collaboratively using SMPC. Experiments show that our approach achieves comparable performance to centralised models while offering significant improvements in privacy and security. This article presents a ground-breaking solution to privacy and security challenges in chatbots, and we hope our approach will foster trust and encourage broader adoption of chatbot technology with privacy at the forefront.