Deep Q-learning based Reinforcement Learning Approach for Radio Frequency Fingerprinting in Internet of Things

Intelligent radio frequency (RF) fingerprinting technology can be used for lightweight device identification by exploiting unique RF characteristics of wireless devices through artificial intelligence (AI). However, in low-power device communication such as bluetooth and lora, it is indeed challenging to improve recognition accuracy and train desired AI models due to limited sample data sets. Besides, it is required to provide a large number of different data to train the model so that the uniqueness and variability of RF characteristics between different wireless devices can be maintained well. To tackle these challenges, we in this paper propose a deep Q-learning-based Reinforcement Learning Approach for RF fingerprint enhancement by integrating the improved reward-driven strategies with the efficient I/Q convolutional network. Particularly, wireless signal features are extracted from raw I/Q data by employing I/Q convolution model. Then a Deep Q Network (DQN) is proposed to realize device recognition. Due to the reinforcement learning approach, the network architecture can be simplified and the convergence of network can be accelerated. Simulation results demonstrate the recognition capabilities of our reinforcement learning-based RF fingerprint recognition method.