[Paper Review] Differential Private Federated Transfer Learning for Mental Health Monitoring in Everyday Settings: A Case Study on Stress Detection

Ziyu Wang, Zhongqi Yang+2 more|arXiv (Cornell University)|Feb 16, 2024

Digital Mental Health Interventions6 citations

TL;DR

The paper proposes a differential private federated transfer learning framework for mental health monitoring, applied to stress detection, integrating pre-training on public data with federated on-device fine-tuning under DP to improve privacy and performance.

ABSTRACT

Mental health conditions, prevalent across various demographics, necessitate efficient monitoring to mitigate their adverse impacts on life quality. The surge in data-driven methodologies for mental health monitoring has underscored the importance of privacy-preserving techniques in handling sensitive health data. Despite strides in federated learning for mental health monitoring, existing approaches struggle with vulnerabilities to certain cyber-attacks and data insufficiency in real-world applications. In this paper, we introduce a differential private federated transfer learning framework for mental health monitoring to enhance data privacy and enrich data sufficiency. To accomplish this, we integrate federated learning with two pivotal elements: (1) differential privacy, achieved by introducing noise into the updates, and (2) transfer learning, employing a pre-trained universal model to adeptly address issues of data imbalance and insufficiency. We evaluate the framework by a case study on stress detection, employing a dataset of physiological and contextual data from a longitudinal study. Our finding show that the proposed approach can attain a 10% boost in accuracy and a 21% enhancement in recall, while ensuring privacy protection.

Motivation & Objective

• Address privacy concerns in mental health monitoring while leveraging limited, diverse data.
• Combine differential privacy with federated learning to protect participant data during model updates.
• Use transfer learning to mitigate data scarcity and imbalance by pre-training on a large public dataset and fine-tuning on-user data.
• Evaluate the framework on a stress detection case study using physiological and contextual data.
• Demonstrate privacy-utility trade-offs and robustness against privacy budget variations.

Proposed method

• Pre-train a universal model on a large public dataset using binary cross-entropy loss.
• Distribute the pre-trained model to clients and fine-tune on user-specific data, applying gradient clipping.
• Add Laplacian noise to gradients with scale Delta f / epsilon to achieve differential privacy during client updates.
• Aggregate noisy client updates via Federated Averaging to form a global model.
• Model uses a three-layer MLP with 12 HR/HRV features derived from PPG and motion data.
• Evaluate privacy-utility trade-offs by varying the privacy budget epsilon.

Figure 1: Overview of the proposed framework.

Experimental results

Research questions

• RQ1 Can differential privacy integrated with federated learning protect sensitive mental health data without severely harming performance?
• RQ2 Does transfer learning via a pre-trained universal model improve performance in data-scarce, personalized stress detection tasks?
• RQ3 What is the impact of the privacy budget (epsilon) on model accuracy, recall, and ROC in this setting?
• RQ4 How does the proposed framework compare to non-federated or non-DP baselines in stress detection?

Key findings

• The integrated framework achieves higher accuracy and recall than baselines: 0.53 accuracy, 0.52 F1, 0.75 recall, 0.40 precision for the fine-tuned model.
• Plain model accuracy 0.43, F1 0.39, recall 0.54, precision 0.31; Pre-trained model accuracy 0.51, F1 0.44, recall 0.58, precision 0.36.
• Differential privacy via Laplacian noise with epsilon = 1 maintains ROC while enabling privacy protections.
• Recall improvement indicates better detection of stressed cases, reducing misses.
• Privacy-utility analysis shows DP-FL does not significantly degrade ROC in reported scenarios.

Figure 2: Overview of the proposed approach.