🤖 AI

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

The paper presents GuidaPA, a privacy-preserving chatbot for the Italian Public Administration that utilizes Federated Learning and QLoRA fine-tuning to achieve high-quality conversational performance on sensitive internal documentation without requiring centralized data sharing.

Original authors: Daniel M. Jimenez-Gutierrez, Albenzio Cirillo, Raffaele Nicolussi, Alessio Beltrame, Andrea Vitaletti

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniel M. Jimenez-Gutierrez, Albenzio Cirillo, Raffaele Nicolussi, Alessio Beltrame, Andrea Vitaletti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of government offices, each sitting in a different building with their own locked filing cabinets full of important rules, manuals, and FAQs. They all want to build a super-smart digital assistant (a chatbot) to help citizens and staff navigate these complex rules.

The problem? The law and privacy rules say they cannot take those filing cabinets out of their buildings and dump them into one giant central warehouse to train the AI. That would be like trying to mix everyone's private recipes into one big pot before cooking; it's too risky and often illegal.

GuidaPA is the solution the researchers built. It's a way to teach the AI without ever moving the filing cabinets.

Here is how it works, using a simple analogy:

The "Secret Recipe" Cooking Class

Think of the AI model as a chef who wants to learn how to cook "Government Stew."

The Old Way (Centralized): Usually, you would gather all the ingredients (the manuals and FAQs) from every office, bring them to one big kitchen, and let the chef taste everything at once. This makes a great stew, but it violates the rule that ingredients must stay in their own kitchens.
The GuidaPA Way (Federated Learning): Instead of moving the ingredients, the chef sends a blank recipe book to every office.
- Office A (Radio Broadcasting) looks at their own manuals, teaches the recipe book a few tricks, and sends the updated book back. They keep their actual manuals locked away.
- Office B (Electronic Communications) does the same with their own manuals, updates the book, and sends it back.
- The Central Server collects all the updated recipe books, mixes the new tricks together to create one "Master Recipe Book," and sends it back out to the offices.

This happens over and over (15 rounds in this study). The chef gets smarter and smarter, learning from all the different offices, but no office ever has to share their actual documents with anyone else.

The "ID Card" Security Guard

The paper also highlights a crucial safety feature called Role-Based Access Control (RBAC).

Imagine the chatbot is a receptionist in a secure building.

If a Citizen walks in asking about Radio Broadcasting rules, the receptionist checks their ID card. The card says, "You can only see the Radio Broadcasting folder." The receptionist answers the question using only that folder.
If that same citizen tries to ask about Electronic Communications, the receptionist says, "Sorry, your ID card doesn't give you access to that folder."
If a Radio Officer walks in, they get access to the Radio folder but not the Electronic folder.

This ensures that even though the AI is smart, it never accidentally spills secrets to the wrong person. It strictly follows the "need to know" rule.

What Did They Find?

The researchers tested this system using real (but public) manuals from two Italian government platforms: SIGESON (for radio broadcasting) and SIDFORS (for electronic communications).

The Result: The chatbot trained this "secret" way (Federated Learning) became almost just as good as the "old way" where all data is combined centrally.
The Proof: They measured how well the AI answered questions using standard scoring systems (like ROUGE and BLEU). The "Secret Recipe" method scored very high, proving you don't need to break privacy rules to get a smart assistant.
The Improvement: Before this training, the AI was like a general encyclopedia—it knew a little about everything but nothing specific about government rules. After the training, it became a true expert in those specific government procedures.

The Bottom Line

GuidaPA shows that government agencies can work together to build a smart, helpful AI assistant without ever having to share their private data. They can keep their documents in their own secure offices, send only "lessons learned" (math updates) to a central hub, and still end up with a highly accurate, secure, and helpful tool for the public.

It's like a group of neighbors learning to bake the perfect pie together by sharing their tips with each other, without ever having to let anyone else into their kitchens to see their secret family recipes.

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

1. Problem Statement

The deployment of Large Language Model (LLM) chatbots within Public Administration (PA) faces significant barriers due to organizational silos, regulatory constraints, and privacy requirements. Conventional centralized training is often impractical because sensitive institutional data (e.g., internal manuals, support tickets, database extracts) cannot be pooled into a single repository without violating data sovereignty or governance policies. Conversely, fully localized solutions are costly and lead to inconsistent behavior across different administrative offices.

The core challenge is to develop a conversational AI system that:

Preserves Data Locality: Keeps sensitive documentation within individual PA offices.
Ensures Domain Adaptation: Adapts general-purpose LLMs to specific PA terminology and procedural workflows (e.g., authorization requests, accounting access).
Enforces Access Control: Guarantees that users only access information relevant to their institutional roles (e.g., citizens vs. officers).
Handles Non-IID Data: Manages the fact that different offices possess related but distinct datasets (non-independent and identically distributed).

2. Methodology

The authors propose GuidaPA, a system architecture that integrates Federated Learning (FL) with role-based access control (RBAC) and secure client-side processing.

System Architecture

GuidaPA consists of four primary components:

Conversation Manager: The entry point for user interaction. It enforces RBAC by validating user credentials and restricting the model's context to authorized data sources. It prevents prompt injection attacks by explicitly blocking access to restricted content based on the user's role (e.g., a SIGESON officer cannot access SIDFORS data).
Federated Learning (FL) Module: Orchestrates collaborative training without sharing raw data. Clients (PA offices) train local models on their private datasets and transmit only model updates (weights) to a central server. The server aggregates these updates using FedAvg (weighted by dataset size) to create a global model, which is then redistributed to clients.
Client Data Pipeline: Handles secure local preprocessing. Raw documents (PDFs/HTML) are extracted, cleaned, and converted into instruction-tuning formats locally. No raw data leaves the client environment.
Performance Module: Monitors global and local model quality and quantifies data heterogeneity (non-IID effects) using statistical measures like Jensen–Shannon Divergence (JSD) and Hellinger Distance (HD).

Experimental Setup

Data Sources: The study utilizes public documentation from two Italian PA platforms managed by the Ministry of Enterprises and Made in Italy (MIMIT):
- SIGESON: Radio broadcasting management (8 pages of manuals + FAQs).
- SIDFORS: Electronic communications network authorization (31 pages of manuals + FAQs).
- Note: These public docs serve as a "safe proxy" for restricted internal sources.
Models: The study evaluates eight LLMs, including Italian-centric models (e.g., cerbero-7b, zefiro-7b) and multilingual baselines (e.g., Llama-3.1-8B, Mistral-7B).
Training Protocol:
- Fine-tuning: Parameter-Efficient Fine-Tuning (PEFT) using QLoRA (4-bit quantization) with LoRA rank $r=32$ .
- Federated Rounds: 15 server rounds with 1 local epoch per round.
- Split: 80/20 train-test split per client.
- Hardware: Single GPU workstation (RTX A6000).

3. Key Contributions

First FL Application in Italian PA: This is the first study to apply Federated Learning to LLM-based chatbots specifically within the Italian Public Administration context, addressing the unique constraints of institutional data silos.
Integrated Deployment Framework: Unlike prior FL research focused solely on model training, GuidaPA integrates application-level mechanisms essential for PA:
- Role-Based Access Control (RBAC) enforced at the conversation level.
- Secure client-side preprocessing pipelines.
- Explicit monitoring of non-IID data effects.
Empirical Validation of Near-Centralized Quality: The study provides evidence that federated fine-tuning can achieve performance comparable to private centralized training while keeping data local.
Quantification of Institutional Non-IID: The authors demonstrate that while PA datasets are non-IID (due to distinct procedural verticals), the divergence is moderate enough to allow stable global aggregation without complex personalization mechanisms.

4. Results

The evaluation compares the Federated Learning (FL) models against a Private Centralized Baseline (where all data is pooled) and a General-Purpose Baseline (pre-trained models without fine-tuning).

Domain Adaptation: Fine-tuning significantly outperforms general-purpose models. For the best model (cerbero-7b-openchat), ROUGE-1 improved from 41.45 to 62.18, and BLEU-4 from 26.97 to 50.90.
Federated vs. Centralized: The best federated model (cerbero-7b-openchat) achieved scores very close to the centralized baseline:
- ROUGE-1: 61.10 (FL) vs. 62.18 (Centralized).
- BLEU-4: 45.02 (FL) vs. 50.90 (Centralized).
- METEOR: 63.94 (FL) vs. 64.17 (Centralized).
Local Performance: Performance varied by client. The cerbero-7b model performed best on the SIGESON dataset, while cerbero-7b-openchat excelled on the SIDFORS dataset, highlighting the impact of data heterogeneity.
Non-IID Analysis: Lexical divergence (measured by HD and JSD) was found to be moderate, consistent with shared regulatory language but distinct domain terminologies, supporting the stability of the federated aggregation.

5. Significance and Claims

The paper claims that GuidaPA demonstrates the feasibility of privacy-preserving conversational AI in public services without the need for centralized data pooling.

Practical Viability: The results suggest that PA entities can collaborate to build high-quality assistants while adhering to strict data sovereignty and governance constraints.
Balanced Trade-offs: The system successfully balances scalability, data protection, and usability. It offers a practical alternative to building centralized data repositories, which are often slow, expensive, or legally constrained.
Governance Support: By integrating RBAC and non-IID monitoring, the framework supports a governance workflow suitable for deployment in environments with evolving local data and strict access policies.

The authors remain modest, noting that the study uses public documentation as a proxy and that the dataset size is small. They frame the results as evidence of feasibility rather than a guarantee that the framework will scale unchanged to larger, noisier, and more heterogeneous institutional corpora without further optimization. Future work is proposed to address stronger non-IID settings, incorporate differential privacy, and include human evaluation.