The paper proposes a method of extracting the feature vector of images, which makes it possible to effectively detect the presence of hidden information in JPEG images embedded by various popular steganography tools. This method is based on the usage of the transition probability matrix. The essence of the method for extracting the feature vector of the image is to use the transition probability matrix and apply the image calibration method to improve the accuracy of steganalysis and reduce the number of false positives. For each image from the training and test sets a feature vector is found in this way, the number of elements is 324. Further, the models were trained on the training dataset by each of machine learning methods separately: decision trees with gradient boosting, linear models, k-nearest neighbors, support vector machines, neural networks, and artificial immune systems. To assess the capacity of the models the following metrics were used: accuracy, the rate of the false positive and false negative errors, and the confusion matrix. The results of classification by each of the above methods are given. For training and testing a dataset IStego100K was used, which consists of 208 thousand images of the same size 1024 x 1024 with different quality values in the range from 75 to 95. One of the J-UNIWARD, nsF5, and UERD steganography algorithms was used to embed a hidden message. As a result, we can observe that the proposed approach to extracting the feature vector makes it possible to detect the presence of hidden information embedded by non-adaptive steganography (Steghide, OutGuess and nsF5) in static JPEG images with high accuracy (more than 95%). However, for adaptive steganography methods (J-UNIWARD, UERD) the accuracy is less (about 50-60%).