Research On Device Identification Of Internet Of Things With Feature Space Shift And Open Category

The rapid spread of Internet of Things(IoT)has brought convenience to life,but it also poses potential security issues.At present,research on IoT security focuses on detection at the application layer,but this cannot solve the attacks like Distributed Denial of Service(DDOS).To prevent attacks that hard to be detected by the application layer,researchers started the research on the security of the physical layer.At the physical layer of wireless devices,due to the influence of manufacturing process,its electronic production elements have fixed physical characteristics.Even if the equipment is produced in the same batch,components such as filters,oscillators,amplifiers,and mixers in the RF module will still have differences.It makes the transmitted wireless signals have unique individual characteristics.The receiver receives the wireless signal,extracts the features of the physical layer,and enters the features signal through the machine learning model.Finally it can tell the signal is transmitted by which equipment.So the unauthorized,forged wireless equipment will be identified,and it can be used to prevent IoT attacks.Although this method provides a good way to solve the IoT attacks,there are still some difficulties.First,the traditional way is to manually extract the features of the physical layer.The biggest difficulty of these methods is that they need to spend a lot of time to design algorithms for extracting characteristic.Meanwhile,its generality is not reasonable,It is necessary to design different algorithms for different types of signals.Second,the characteristics of the device may shift slightly when the hardware of the equipment ages or the wireless environment changes.The model trained before is no longer suitable for future target samples,since the individual identification is based on characteristics-stabilized,which brings researchers difficulties.Besides,all types of samples cannot be collected in the environment,resulting in the incomplete training set.In the supervised learning or semi-supervised learning,each test sample will be judged as one of the known classes.Due to the training set is incomplete,the unseen class may be incorrectly judged as a known class.This paper adopts deep learning,transfer learning algorithm,and generative adversarial networks(GAN)algorithm to solve the above three problems.The experimental results show that deep learning can achieve higher recognition rate under very complex environment.It can achieve a 94%recognition rate.The transfer learning can be improved by about 10%with very few target samples.And the open-category classification based on GAN can achieve a 90%accuracy rate with a recall rate of 97%.It is verified that the improved method proposed in this paper is effective.