With the accelerating digital transformation of power systems

the trend toward open data sharing has intensified the severity of data inference threats. This paper focuses on the security of power flow information

a key sensitive data in power systems. First

an inference model for power flow information is constructed

revealing the nonlinear relationship between public data and sensitive power flow data through a multi-level coupling relationship framework. Based on this

a data-mechanism hybrid-driven power flow information inference method is proposed

which informsphysical constraints of power systems into a deep graph learning approach to achieve accurate power flow information inference that conforms to power systems operation rules. Comparative experiments show that this method has superior inference performance compared to traditional neural network methods. The research further analyzes the influence mechanism of multi-source data synergy on information leakage risks

and systematically evaluates the protective effects of typical data protection techniques. An optimized defense strategy that balances data security and usability is proposed

providing methodological support for data security in power systems.