In LoRaWAN networks, devices are identified by two identifiers: a globally unique and stable one called DevEUI, and an ephemeral and randomly assigned pseudonym called DevAddr. The association between those identifiers is only known by the network and join servers, and is not available to a passive eavesdropper.In this work, we consider the problem of linking the DevAddr with the corresponding DevAddr based on passive observation of the LoRa traffic transmitted over the air. Leveraging metadata exposed in LoRa frames, we devise a technique to link two messages containing respectively the DevEUI and the DevAddr, thus identifying the link between those identifiers. The approach is based on machine learning algorithms using various pieces of information including timing, signal strength, and fields of the frames. Based on an evaluation using a real-world dataset of 11 million messages, with ground truth available, we show that multiple machine learning models are able to reliably link those identifiers. The best of them achieves an impressive true positive rate of over 0.8 and a false positive rate of 0.001.