Internet of Things systems are growing rapidly, offering services in diverse environments but also posing important security challenges. Among the most common threats are denial of service attacks, which compromise the availability of resource-constrained devices. One of the most widely used protocols in these settings is the Constrained Application Protocol, whose features can be exploited to amplify such attacks. In this work, we compare three unsupervised techniques for reducing dimensionality: Isometric Mapping, t-distributed stochastic neighbor embedding, and Uniform Manifold Approximation and Projection, for visual analysis, detection, and characterization of denial of service attacks on networks using that protocol. We use a real traffic dataset obtained in an experimental setup under denial of service attack conditions. The results show that t-distributed stochastic neighbor embedding and Uniform Manifold Approximation and Projection enable clear visual separation between legitimate and malicious traffic. These findings support the use of low-dimensional latent representations as an exploratory tool to detect anomalous behavior in Internet of Things networks.