Learning-based image steganalysis is an effective and universal approach to cope with the following two difficulties: unknown statistics and steganographic algorithms. A crucial part of the learning-based process is the selection of low-dimensional features, which strongly impacts the accuracy of classification. A novel principal feature selection and fusion (PFSF) method is presented to reduce features, and then it is applied to image steganalysis. First, we analyze the multicollinearity among features to eliminate redundant features. Next, we implement the linear transform based on principal components analysis (PCA) and use Savage decision-making to eliminate insignificant features. Last, in order to further reduce features, we fuse the selected features, followed by selecting the principal features from the fused features to form a new feature set. The advantage of the proposed method is that it needs the cover images only, without requiring the availability of the stego-images in the process of the features selection. Moreover, the proposed method greatly reduces the computational time. Our method has been tested on two feature sets from Moulin's and Fridrich's features. The experimental results show that our method not only reduces the feature number by 90%, but also provides more reliable detection results than the previous steganalysis methods do.