Deep Transfer Learning for Physical Layer Security in Wireless Communication Systems

Machine learning has been used in wireless networking for wireless RF allocation, power allocations, predicting traffic load and inferencing/estimating channels. However, machine learning in wireless networking faces several challenges because of the big or small data, unlabeled data, system complexity, long training process, resource constrained (energy, storage, computing) wireless devices, and constantly changing user mobility, dynamic network topology, susceptibility of wireless jammers and eavesdroppers, and operating wireless environment. In order to address these challenges, transfer learning is regarded as one of the viable solutions as it leverages the knowledge from similar past tasks for learning of new problems. In this paper, we investigate deep transfer learning for physical layer security in wireless networks with large-scale multiple-input multiple output (MIMO) system where previously trained model is transferred to a new wireless environment to predict secure waveform vectors and power levels for users who coexist with eavesdroppers (who overhear communications passively for resulting in lower secrecy rate for legitimate users) and jammers (who actively jam the channels or deteriorate the received signal at the legitimate receivers). In this case, a legitimate transmitter adapts it's transmit waveform and power using transfer learning to avoid jamming and eavesdropping impacts. This is analogous to training war-fighters in a training camp and deploying them to a completely new real battlefield where they are expected to (learn instantly and) defeat the enemy based on their past training. We evaluate the performance of the proposed approach using numerical results obtained from extensive simulations.