Privacy Preserving Topic-wise Sentiment Analysis of the Iran Israel USA Conflict Using Federated Transformer Models

Privacy protection through federated learning: The study demonstrates how sentiment analysis can be conducted without centralizing sensitive user data. By using federated learning, the researchers trained AI models across distributed systems while keeping individual YouTube comments on local devices, addressing a key concern in responsible AI about data privacy.

Human validation improves AI reliability: The researchers initially used an automated sentiment analyzer (VADER) but then had humans manually inspect the results to validate accuracy. This hybrid approach highlights the importance of human oversight in AI systems, particularly when analyzing nuanced topics like geopolitical conflicts where automated tools alone may misclassify sentiment.

Explainability makes AI decisions transparent: The study applied SHAP (an Explainable AI technique) to interpret why the sentiment models made specific predictions. This transparency is crucial for responsible AI, especially in sensitive applications like conflict analysis where stakeholders need to understand and trust how conclusions were reached.

Multiple AI models tested for robustness: Rather than relying on a single algorithm, the researchers compared six different transformer-based models (BERT, RoBERTa, XLNet, DistilBERT, ModernBERT, and ELECTRA). Testing multiple approaches helps identify which models are most reliable and reduces the risk of deploying flawed systems for public opinion analysis.

Topic-specific analysis provides context: By using LDA to identify distinct discussion topics before analyzing sentiment, the study avoids treating all comments equally. This nuanced approach recognizes that public opinion on complex geopolitical issues varies by specific aspect, leading to more responsible and accurate interpretation of global sentiment.