Attacks against Supervisory Control and Data Acquisition (SCADA) systems operating critical infrastructures have largely appeared in the past decades. There are several anomaly detection systems that model the traffic of request-response mechanisms, where a client initiates a request to a server and the server sends back a response later. However, many modern SCADA protocols also allow server-driven traffic without a paired request, and anomaly detection for server-driven traffic has not been well-studied. This paper provides a comprehensive understanding of server-driven traffic across different protocols, such as MMS, Siemens S7, S7-plus, and IEC 60870-5-104 (IEC-104), with traffic analysis. The analysis results show that the common postulation of periodicity and correlation within SCADA traffic holds true for most of the analyzed datasets. The paper then proposes a Multivariate Correlation Anomaly Detection (MCAD) approach for server-driven traffic that presents complicated correlations among flows. The proposed approach is compared with a univariate correlation anomaly detection approach designed for SCADA and a general purpose anomaly detection approach based on neural network techniques. These approaches are tested with an IEC-104 dataset from a real power utility with injected timing perturbations resulting from a Stuxnet-like stealthy attack scenario. The detection accuracy of MCAD outperforms the compared methods and the time-to-detection performance is promising.