This research introduces FedTransient, a novel federated learning framework enabling collaborative training of transient classifiers across multiple observatories without centralizing sensitive data. It addresses instrument heterogeneity, statistical heterogeneity, and privacy-preserving knowledge aggregation, achieving 87.3% classification accuracy across heterogeneous instruments.
Key findings
FedTransient enables collaborative training of transient classifiers across observatories without data centralization.
The framework addresses instrument heterogeneity and statistical heterogeneity from non-IID class distributions.
Privacy-preserving knowledge aggregation is achieved, maintaining data privacy.
The proposed framework achieves 87.3% classification accuracy across heterogeneous instruments.
The work establishes a methodological foundation for collaborative transient classification in the multi-survey era.
Limitations & open questions
The framework's scalability to a larger number of observatories and longer time-series data needs further investigation.
The real-world application and integration with existing observatory infrastructures pose additional challenges.