Collaborative recommender systems are vulnerable to recommendation attack, in which malicious users insert fake profiles into the rating database in order to bias the systems output. To reduce this risk, many methods have been proposed to detect such attack. Despite their effectiveness, a lot of these methods are built based on the hand-designed features which are usually difficult to extract even for domain experts. In order to build the detection method without resorting to hand-designed features, in this paper, we propose a deep learning-based approach for detecting recommendation attack (called DL-DRA). The proposed approach can learn directly from the low-level rating data for the training of classifier. Therefore, it does not have the problem of how to extract hand-designed features. We first propose a framework to show the basic structure of the proposed detection approach. Then, we propose a rating matrix generation method to transform the rating vector into rating matrix for each user. After that, we use the bicubic interpolation algorithm to resize the rating matrix in order to reduce the sparsity of the rating matrix. Finally, on the basis of convolutional neural network (CNN), we construct a deep learning network which can learn directly from the resized rating matrix. On this basis, we propose an algorithm for detecting the recommendation attack. We conduct a large number of comparative experiments with the state-of-the-art methods for recommendation attack on different scale MovieLens datasets. The experimental results show that the proposed approach can detect the recommendation attack, effectively and steadily.