This paper presents a method for steganalysis of images based on principal component analysis and support vector machine. Considering that embedding information inevitably brings noise to cover images, we design and calculate eigenvectors of images after analyzing principal components of image blocks. By using eigenvectors of cover images and steg-images to train SVM, we can obtain a trained SVM to be used to blind detection. Experimental results show that the proposed scheme is effective to many steganographic methods and correct detection rate is better than the steganalysis method using image quality metrics.