Distributed secrecy for information theoretic sensor network models

This dissertation presents a novel problem inspired by the characteristics of sensor networks. The basic setup through-out the dissertation is that a set of sensor nodes encipher their data without collaboration and without any prior shared secret materials. The challenge is dealt by an eavesdropper who intercepts a subset of the enciphered data and wishes to gain knowledge of the uncoded data. This problem is challenging and novel given that the eavesdropper is assumed to know everything, including secret cryptographic keys used by both the encoders and decoders. We study the above problem using information theoretic models as a necessary first step towards an understanding of the characteristics of this system problem. This dissertation contains four parts. The first part deals with noiseless channels, and the goal is for sensor nodes to both source code and encipher their data. We derive inner and outer regions of the capacity region (i.e the set of all source coding and equivocation rates) for this problem under general distortion constraints. The main conclusion in this part is that unconditional secrecy is unachievable unless the distortion is maximal, rendering the data useless. In the second part we thus provide a practical coding scheme based on distributed source coding using syndromes (DISCUS) that provides secrecy beyond the equivocation measure, i.e. secrecy on each symbol in the message. The third part deals with discrete memoryless channels, and the goal is for sensor nodes to both channel code and encipher their data. We derive inner and outer regions to the secrecy capacity region, i.e. the set of all channel coding rates that achieve (weak) unconditional secrecy. The main conclusion in this part is that interference allows (weak) unconditional secrecy to be achieved in contrast with the first part of this dissertation. The fourth part deals with wireless channels with fading and additive Gaussian noise. We derive a general outer region and an inner region based on an equal SNR assumption, and show that the two are partially tight when the maximum available user powers are admissible.