Privacy Preserving Topic-wise Sentiment Analysis of the Iran Israel USA Conflict Using Federated Transformer Models
A new AI framework analyzes YouTube comments on geopolitical conflicts while keeping user data private.
A team of researchers has published a novel framework for conducting privacy-preserving sentiment analysis on a highly sensitive topic: the Iran-Israel-USA conflict. By collecting and preprocessing approximately 19,000 user comments from major international YouTube news channels, the study applies Latent Dirichlet Allocation (LDA) to identify key discussion topics. The core technical achievement is the fine-tuning of several transformer-based models—including BERT, RoBERTa, and ELECTRA—for sentiment classification, with ELECTRA emerging as the top performer.
To address the critical issue of data privacy, the researchers integrated the best-performing model into a federated learning environment. This approach allows the AI model to be trained across multiple decentralized devices or servers holding local data samples, without exchanging the raw data itself. In a two-client configuration, the federated model maintained a high accuracy of 89.59%. Furthermore, the team applied Explainable AI (XAI) techniques using SHAP to interpret the model's predictions, identifying the specific words that most influenced sentiment scores and adding a layer of transparency to the analysis.
- ELECTRA model achieved 91.32% accuracy in classifying sentiment from 19,000 YouTube comments on the conflict.
- Federated learning implementation preserved privacy with only a 1.73% accuracy drop, achieving 89.59% in a distributed setup.
- Framework combines topic modeling (LDA), transformer fine-tuning, and explainable AI (SHAP) for a complete, privacy-conscious analysis pipeline.
Why It Matters
Enables organizations and researchers to gauge public opinion on sensitive issues without compromising individual privacy or centralizing data.