Estimation of conditional average treatment effects on distributed data: A privacy-preserving approach