Nowadays, more and more Internet-of-Things (IoT) smart products, interconnected through various wireless communication technologies (Wifi, Bluetooth, Zigbee, Z-wave, etc.) are integrated in daily life, especially in homes, factories, cities, etc. Such IoT technologies have become very attractive with a large variety of new services offered to improve the quality of life of the endusers or to create new economic markets. However, the security of such connected objects is a real concern due to weak or flawed security designs, configuration errors or imperfect maintenance. Moreover, the vulnerabilities discovered in IoT products are often difficult to eliminate because, most of the time, they cannot be patched easily. Therefore, protection mechanisms are needed to mitigate the potential risks induced by such objects in private and public connected areas. In this paper, we propose a novel approach to detect potential attacks in smart places (e.g. smart homes) by detecting deviations from legitimate communication behavior, in particular at the physical layer. The proposed solution is based on the profiling and monitoring of the Radio Signal Strenght Indication (RSSI) associated to the wireless transmissions of the connected objects. A machine learning neural network algorithm is used to characterize legitimate communications and to identify suspiscious scenarios. We show the feasibility of this approach and discuss some possible application cases.