Image steganography delivers secret data by slight modifications of the cover. To detect these data, steganalysis tries to create some features to embody the discrepancy between the cover and steganographic images. Therefore, the urgent problem is how to design an effective classification architecture for given feature vectors extracted from the images. We propose an approach to automatically select effective features based on the well-known JPEG steganographic methods. This approach, referred to as extreme learning machine revisited feature selection (ELM-RFS), can tune input weights in terms of the importance of input features. This idea is derived from cross-validation learning and one-dimensional (1-D) search. While updating input weights, we seek the energy decreasing direction using the leave-one-out (LOO) selection. Furthermore, we optimize the 1-D energy function instead of directly discarding the least significant feature. Since recent Liu features can gain considerable low detection errors compared to a previous JPEG steganalysis, the experimental results demonstrate that the new approach results in less classification error than other classifiers such as SVM, Kodovsky ensemble classifier, direct ELM-LOO learning, kernel ELM, and conventional ELM in Liu features. Furthermore, ELM-RFS achieves a similar performance with a deep Boltzmann machine using less training time.