GuidaPA: Privacy-Preserving Chatbot for Public Administration via Federated Learning

· Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Advanced, quick

Summary

GuidaPA is a privacy-preserving chatbot developed for the Italian Public Administration (PA), trained using Federated Learning (FL) on public documentation from SIGESON (8 pages) and SIDFORS (31 pages) platforms. This system is designed for future deployment with restricted internal data, avoiding central pooling due to regulatory and organizational constraints. GuidaPA incorporates role-based access control, secure client-side preprocessing, explicit monitoring of non-IID effects, and parameter-efficient federated fine-tuning via QLoRA (4-bit). Evaluated over 15 federated rounds with an 80/20 train-test split, the best federated model achieved ROUGE-1/2/L scores of 61.10/55.77/59.44, BLEU-4 of 45.02, and METEOR of 63.94. These results are comparable to private centralized fine-tuning, demonstrating FL's capability to deliver high-quality conversational AI for public services without centralizing sensitive data. Domain fine-tuning significantly improved ROUGE-1 from 41.45 to 62.18 and BLEU-4 from 26.97 to 50.90 compared to a general-purpose baseline.

Key takeaway

For AI Architects and ML Engineers deploying conversational AI in highly regulated sectors like public administration, GuidaPA demonstrates that Federated Learning is a robust solution. You can achieve high-quality chatbot performance comparable to centralized methods while strictly adhering to data privacy and localization requirements. Consider implementing FL with parameter-efficient fine-tuning for your next project involving sensitive, distributed data sources.

Key insights

Federated Learning enables high-quality, privacy-preserving chatbots for sensitive public administration data without centralizing information.

Principles

Method

GuidaPA uses QLoRA (4-bit) for parameter-efficient federated fine-tuning over 15 rounds with an 80/20 train-test split per client, explicitly monitoring non-IID effects.

In practice

Topics

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.