Comparison of Machine Learning Approaches Based on Multiple Channel Attributes for Authentication and Spoofing Detection at the Physical Layer

The aim of this study is to assess the effectiveness of Physical Layer Authentication (PLA) in securing IoT nodes. Specifically, we present a PLA framework based on wireless fingerprinting, where the legitimated node is distinguished from potential attackers by exploiting the unique wireless channel features. To achieve this objective, we employ various machine learning approaches for anomaly detection, making use of a wide range of channel attributes in time-varying conditions. In particular, four different Machine Learning (ML) strategies in their one class version have been considered and compared: decision-tree, kernelbased, clustering and nearest neighbors. Our study highlights advantages and disadvantages of each method, considering parameters optimization, training requirements and time complexity. Results show that the use of multiple-attribute allows to achieve accurate detection performance. In particular, our results reveal that the kernel-based solution is the one that achieves best results in terms of accuracy, but the nearest neighbor’s solution has very similar performance with a significant advantage in terms of complexity and no need for training, making it more suitable for time-varying contexts, and a promising choice for securing IoT nodes through PLA based on wireless fingerprinting. The other two alternatives have somewhat lower performance but low complexity. This research contributes valuable insights into enhancing IoT security through PLA techniques.