[Paper Review] FedBot: Enhancing Privacy in Chatbots with Federated Learning

Addi Ait‐Mlouk, Sadi Alawadi+2 more|arXiv (Cornell University)|Apr 4, 2023

Privacy, Security, and Data Protection4 citations

TL;DR

FedBot proposes a privacy-preserving customer support chatbot that leverages federated learning (FL) to train deep bidirectional Transformer models without centralizing sensitive user data. By enabling collaborative, decentralized training with incremental learning, FedBot achieves strong performance comparable to centralized models while ensuring data privacy, scalability, and compliance with regulations like GDPR.

ABSTRACT

Chatbots are mainly data-driven and usually based on utterances that might be sensitive. However, training deep learning models on shared data can violate user privacy. Such issues have commonly existed in chatbots since their inception. In the literature, there have been many approaches to deal with privacy, such as differential privacy and secure multi-party computation, but most of them need to have access to users' data. In this context, Federated Learning (FL) aims to protect data privacy through distributed learning methods that keep the data in its location. This paper presents Fedbot, a proof-of-concept (POC) privacy-preserving chatbot that leverages large-scale customer support data. The POC combines Deep Bidirectional Transformer models and federated learning algorithms to protect customer data privacy during collaborative model training. The results of the proof-of-concept showcase the potential for privacy-preserving chatbots to transform the customer support industry by delivering personalized and efficient customer service that meets data privacy regulations and legal requirements. Furthermore, the system is specifically designed to improve its performance and accuracy over time by leveraging its ability to learn from previous interactions.

Motivation & Objective

• Address the growing concern of data privacy in data-driven chatbots, especially in customer support where sensitive user interactions are common.
• Overcome limitations of centralized training, which risks data breaches and violates privacy regulations like GDPR and HIPAA.
• Develop a proof-of-concept system that enables collaborative model training across multiple clients while keeping all training data localized on user devices.
• Enhance model accuracy and adaptability over time through incremental online learning without compromising user data privacy.
• Demonstrate the feasibility of applying federated learning to large-scale NLP tasks such as dialogue understanding in real-world customer service scenarios.

Proposed method

• Employed a federated learning framework where local clients (e.g., customer support systems) train a shared global model using their private data without uploading raw data to a central server.
• Integrated Deep Bidirectional Transformer models (e.g., BERT-like architectures) for natural language understanding in dialogue systems.
• Used a central aggregation server (combiner) to periodically update the global model by averaging local model updates from participating clients.
• Implemented incremental online learning to allow clients to continuously improve the model with new data, enhancing long-term performance and adaptability.
• Designed the system to support both cross-silo (organization-level) and cross-device (end-user device) FL settings for broad deployment flexibility.
• Optimized FedBot for low-resource environments, enabling deployment on standard laptops with minimal hardware (2 cores, 2GB RAM), ensuring accessibility and scalability.

Figure 1: The Fedbot architecture is organized in three main logical layers, the first one deals with privacy-preserving training using federated learning, the second one is a dialogue manager, and the third is for incremental learning and private data labelling.

Experimental results

Research questions

• RQ1 Can federated learning effectively train high-performing dialogue models for customer support chatbots without centralizing sensitive user data?
• RQ2 How does the performance of a federatedly trained Transformer model compare to a centralized baseline model in terms of accuracy and response relevance?
• RQ3 To what extent does incremental learning in FedBot improve model accuracy and adaptability over time with new local data?
• RQ4 Can FedBot maintain strong privacy guarantees while scaling across diverse, geographically distributed clients in real-world customer support environments?
• RQ5 How does the integration of federated learning with large-scale NLP models impact training efficiency, communication overhead, and system robustness?

Key findings

• FedBot achieved performance comparable to a centralized baseline model after 20 training rounds on customer support datasets, demonstrating that federated training can yield high-quality dialogue models without data centralization.
• The system successfully preserved data privacy by ensuring that no raw user utterances were ever transferred to a central server, aligning with GDPR and other privacy regulations.
• Incremental learning enabled continuous model improvement as new data was added, allowing the model to adapt to evolving user queries and improve over time without retraining from scratch.
• FedBot demonstrated scalability and low resource requirements, functioning effectively on standard laptops with minimal hardware (2 cores, 2GB RAM), making it deployable across diverse client environments.
• The integration of federated learning with deep Transformer models enabled robust, accurate, and privacy-preserving dialogue systems suitable for production use in customer support.
• The framework supported both cross-silo and cross-device FL settings, enabling collaboration across organizations and end-user devices while maintaining data locality and trust.

Figure 2: The Transformer model architecture Vaswani et al. ( 2017 ) - Adapting the transformer model to a Fedbot architecture for natural language processing, featuring an encoder-decoder architecture that generates responses to user inputs using self-attention and feed-forward neural networks