Encryption has became an indispensable technique for image/video-based applications. This has led to the development of many image encryption algorithms, such as perceptual/selective encryption methods which represent an effective way for the security and confidentiality of images. However, few studies focus on visual security metric, which is very important tool for evaluating the effectiveness of these encryption methods. Most of the adopted metrics are the classical randomness-based measures or the objective image quality assessment metrics. However, these metrics showed their limits as a visual security metric, because they do not deal with the content intelligibility, which is one of the key security requirements. Consequently, in this paper, we propose a no-reference (NR) visual security metric for perceptually encrypted images based on multi-output learning called VSMML. The proposed metric consists of a convolutional neural network (CNN) taking as input an encrypted image and providing two outputs corresponding to the visual security (VS) and visual quality (VQ) scores. Experiments were performed on two publicly perceptually encrypted image databases and the results show that the proposed metric significantly outperforms the state-of-the-art methods for visual security and quality assessment tasks. The source code is available at: https://github.com/Mamadou-Keita/VSMML.