Machine Learning for Physical Layer Security

点击翻译按钮获取中文摘要

The digitalization of information processing disruptively changes everyone's life by making information available almost everywhere at any time. With this comes the need of spectrally efficient (wireless) communication systems and, in particular, sophisticated security mechanisms that secure the communication against adversarial attacks and protect the privacy of the data and users.Security related tasks are currently realized on higher layers and usually based on cryptographic principles. These have a wide variety of use and are based on the assumption of insufficient computational capabilities of adversaries and computational hardness of certain problems. However, due to increasing computational power, improved algorithms, and recent advances in number theory, these approaches are becoming less and less secure. Recently, the concept of physical layer security or information theoretic security has been examined as a complement to cryptographic techniques. Such approaches establish reliable communication and unconditional security jointly at the physical layer by exploiting physical properties of the communication channel. However, practical implementations are still in its infancy due to challenges such as its generalizability to arbitrary and changing network configurations and channel conditions.In another line of work, it has been demonstrated that fast and reliable communication schemes can be learned by communication systems using machine learning tools; particularly by using so-called deep neural networks (deep learning). Such machine learning tools can help to solve some of the challenges the communication theory is faced with and provide a way to design sophisticated communication systems that do not need to be tuned by hand to specific channel conditions but are flexible and applicable to a broad range of scenarios. The present proposal tackles the challenge of developing machine learning based echniques for secure (wireless) communication systems. The core of the proposal are three key points:The first goal will be to identify, investigate, and to develop suitable security metrics. Such metrics need to be chosen and designed such that they keep their operational meaning of security and further allow the incorporation into training of different learning algorithms. The second goal will be to develop physical layer security models within the deep learning framework. This part will embrace recent developments for coding and communication and expand those models for physical layer security. The third goal will be to look into more general deep learning concepts including reinforcement learning, recurrent neural networks and generative adversarial networks and efficiently implement the resulting techniques and algorithms for real world scenarios.