Smart cities work with large volumes of data from sensor networks and other sources. To prevent data from being compromised by attacks or errors, smart city IT administrators need to apply attack detection techniques to evaluate possible incidents as quickly as possible. Machine learning has proven to be effective in many fields and, in the context of wireless sensor networks (WSNs), it has proven adequate to detect attacks. However, a smart city poses a much more complex scenario than a WSN, and it has to be evaluated whether these techniques are equally valid and effective. In this work, we evaluate two machine learning algorithms (support vector machines (SVM) and isolation forests) to detect anomalies in a laboratory that reproduces a real smart city use case with heterogeneous devices, algorithms, protocols, and network configurations. The experience has allowed us to show that, although these techniques are of great value for smart cities, additional considerations must be taken into account to effectively detect attacks. Thus, through this empiric analysis, we point out broader challenges and difficulties of using machine learning in this context, both for the technical complexity of the systems, and for the technical difficulty of configuring and implementing them in such environments.