A Personalized and Privacy-Preserving Federated Transformer Framework for Multilingual Sentiment Analysis.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Multilingual sentiment analysis faces challenges with diverse languages and data privacy.
• Federated learning (FL) offers a privacy-preserving approach but struggles with personalization and non-IID data.

Purpose of the Study:

• To introduce FedPerX, a novel federated transformer framework for personalized and privacy-preserving multilingual sentiment analysis.
• To address linguistic heterogeneity, non-IID data, and strict privacy requirements in FL settings.

Main Methods:

• FedPerX utilizes a frozen multilingual backbone (XLM-R) with lightweight, client-specific adapters for personalization.
• Adaptive multi-granular differential privacy is implemented via dynamic noise injection at feature and adapter levels.
• The framework is evaluated on MARC and TSMD multilingual benchmarks across various languages and data types.

Main Results:

• FedPerX achieved consistent improvements over state-of-the-art baselines, with up to +4.3% gains in macro-F1.
• Demonstrated a 70% reduction in communication overhead and minimal variance in client performance.
• Analyses confirmed robustness, adaptability, fairness, and an effective privacy-utility trade-off.

Conclusions:

• FedPerX offers a scalable, personalized, and privacy-preserving solution for federated multilingual sentiment analysis.
• The framework effectively balances model performance, communication efficiency, and user privacy.