Internet of Things (IoT) has emerged as a cutting-edge technology that is changing human life. The rapid and widespread applications of IoT, however, make cyberspace more vulnerable, especially to IoT-based attacks in which IoT devices are used to launch attack on cyber-physical systems. Given a massive number of IoT devices (in order of billions), detecting and preventing these IoT-based attacks are critical. However, this task is very challenging due to the limited energy and computing capabilities of IoT devices and the continuous and fast evolution of attackers. Among IoT-based attacks, unknown ones are far more devastating as these attacks could surpass most of the current security systems and it takes time to detect them and "cure" the systems. To effectively detect new/unknown attacks, in this article, we propose a novel representation learning method to better predictively "describe" unknown attacks, facilitating supervised learning-based anomaly detection methods. Specifically, we develop three regularized versions of autoencoders (AEs) to learn a latent representation from the input data. The bottleneck layers of these regularized AEs trained in a supervised manner using normal data and known IoT attacks will then be used as the new input features for classification algorithms. We carry out extensive experiments on nine recent IoT datasets to evaluate the performance of the proposed models. The experimental results demonstrate that the new latent representation can significantly enhance the performance of supervised learning methods in detecting unknown IoT attacks. We also conduct experiments to investigate the characteristics of the proposed models and the influence of hyperparameters on their performance. The running time of these models is about 1.3 ms that is pragmatic for most applications.