An important issue concerning real-world deployment of steganalysis systems is the computational cost of acquiring features used in building steganalyzers. Conventional approach to steganalyzer design crucially assumes that all features required for steganalysis have to be computed in advance. However, as the number of features used by typical steganalyzers grow into thousands and timing constraints are imposed on how fast a decision has to be made, this approach becomes impractical. To address this problem, we focus on machine learning aspect of steganalyzer design and introduce a decision tree based approach to steganalysis. The proposed steganalyzer system can minimize the average computational cost for making a steganalysis decision while still maintaining the detection accuracy. To demonstrate the potential of this approach, a series of experiments are performed on well known steganography and steganalysis techniques.