Traditional machine learning architectures require data to be centrally stored before model training, which is easy to implement and remains the mainstream application architecture in the industry. However, with the growing capabilities of big data analysis, privacy issues caused by centralized data collection have become more prominent. On one hand, untrustworthy data collectors may leak user data either actively or passively. On the other hand, once collected, user data is no longer under their control, making it difficult to trace and hold accountable in case of a data breach.