TLDR: FedMentor is a new framework for fine-tuning large language models (LLMs) in sensitive areas like mental health. It uses federated learning, low-rank adaptation (LoRA), and domain-aware differential privacy to ensure strong data confidentiality while maintaining model performance and safety. Each client applies privacy noise based on its data’s sensitivity, and the central server adjusts noise levels to balance privacy and utility. This approach significantly improves safety and reduces toxicity in LLM outputs, with minimal impact on model accuracy, making it practical for secure mental health AI deployments.
Large Language Models (LLMs) are increasingly being explored for their potential to offer scalable support in mental health. However, deploying these powerful AI tools in such a sensitive domain comes with significant challenges, primarily concerning user privacy and data confidentiality. Regulations like HIPAA and GDPR impose strict requirements, making traditional centralized training methods difficult due to the need to aggregate highly sensitive user data.
A new framework called FedMentor has been proposed to address these critical issues. Developed by Nobin Sarwar and Shubhashis Roy Dipta from the University of Maryland Baltimore County, FedMentor is designed to enable the privacy-preserving adaptation of LLMs for mental health applications. The core idea is to balance strict confidentiality with the model’s usefulness and safety.
FedMentor integrates three key technologies: Federated Learning (FL), Low-Rank Adaptation (LoRA), and domain-aware Differential Privacy (DP). Federated Learning allows LLMs to be fine-tuned collaboratively across multiple clients (like different clinics or individual devices) without ever centralizing the raw, sensitive user data. Instead of sharing data, clients only share model updates.
To make this process efficient, FedMentor uses Low-Rank Adaptation (LoRA). LoRA is a technique that allows for the fine-tuning of LLMs by only updating a small fraction of the model’s parameters, known as adapters. This significantly reduces the amount of data that needs to be communicated between clients and the central server, making the process much more practical for resource-constrained environments like single-GPU clients.
The most innovative aspect of FedMentor is its domain-aware Differential Privacy. Differential Privacy is a strong mathematical guarantee of privacy, ensuring that individual data points cannot be inferred from the aggregated model updates. FedMentor takes this a step further by assigning a custom privacy budget to each client (or domain) based on the sensitivity of its data. For instance, data related to interpersonal risk factors might receive a stricter privacy budget (meaning more noise is added to its updates) than less sensitive data. The central server also adaptively reduces this privacy noise if the model’s performance (utility) falls below a certain threshold, creating a dynamic balance between privacy and model effectiveness.
Experiments conducted on three mental health datasets (Dreaddit, IRF, and MultiWD) using various LLM backbones demonstrated FedMentor’s effectiveness. The framework significantly improved safety, leading to higher rates of safe outputs and reduced toxicity, compared to standard Federated Learning without privacy. Crucially, it achieved this while maintaining model utility (measured by metrics like BERTScore F1 and ROUGE-L) within a very small margin (0.5%) of the non-private baseline and close to the performance of a hypothetically centralized model.
From an efficiency standpoint, FedMentor proved highly practical. It required less than 173MB of communication per round for models with up to 1.7 billion parameters, a drastic reduction compared to the gigabytes required for full model updates. This low communication overhead and memory footprint make it feasible for deployment on single-GPU clients, which is vital for real-world healthcare settings.
Ablation studies further highlighted the importance of FedMentor’s design choices. Removing domain-specific privacy budgets or adaptive noise control led to lower accuracy and increased disparities among clients, underscoring the value of these integrated features.
Also Read:
- AI Framework Empathy-R1 Offers Deeper Mental Health Support
- Unmasking Hidden Privacy Risks in AI Collaboration
In conclusion, FedMentor offers a robust and practical solution for fine-tuning LLMs in sensitive domains like mental health. By combining federated learning, efficient LoRA adapters, and intelligent domain-aware differential privacy, it paves the way for safer and more trustworthy AI deployments in healthcare and other fields where confidentiality is paramount. You can read the full research paper here: FedMentor: Domain-Aware Differential Privacy for Heterogeneous Federated LLMs in Mental Health.
