In the last decades, Internet of Things objects have been increasingly integrated into smart environments. Nevertheless, new issues emerge due to numerous reasons such as fraudulent attacks, inconsistent sensor behaviours, and network congestion. These anomalies can have a drastic impact on the global Quality of Service in the Local Area Network. Consequently, contextual anomaly detection using network traffic metadata has received a growing interest among the scientific community. The detection of temporal anomalies helps network administrators anticipate and prevent such failures. In this paper, we propose RESIST, a Robust transformEr developed for unSupervised tIme Series anomaly deTection. We introduce a robust learning strategy that trains a Transformer to model the nominal behaviour of the network activity. Unlike competing methods, our approach does not require the availability of an anomalyfree training subset. Relying on a contrastive learning-based robust loss function, RESIST automatically downweights atypical corrupted training data, to reduce their impact on the training optimization. Experiments on the CICIDS17 public benchmark dataset show an improved accuracy of our proposal in comparison to recent state-of-the-art methods.