ESE PhD Thesis Defense: “Learning, Privacy, and Reliable Communication in Large Data Networks”

This thesis explores advancements in three distinct domains: communications, privacy, and machine learning. Within the realm of communication, a comprehensive study is conducted on channel coding at low capacity, a critical aspect of Internet of Things (IoT) technology requiring reliable transmission over channels with minimal capacity. Despite existing finite-length analyses yielding inaccurate predictions and current coding schemes proving inefficient in low-capacity scenarios, this thesis addresses these limitations. It characterizes the finite-length fundamental limits of channel coding for essential channels, offering provably efficient code designs tailored for low-capacity environments. In the domain of privacy, attention is directed towards a decentralized consensus problem handled in a private manner. Existing methods for classical consensus problems often involve openly exchanging private information. This thesis proposes an algorithmic framework capable of achieving the exact limit and the fastest possible convergence rate while safeguarding the privacy of users’ local values. Additionally, a novel information-theoretic metric is introduced to effectively measure the privacy of a node concerning another node within the network. In the machine learning domain, a rigorous mathematical framework is established to investigate the function class of Graph Neural Networks (GNNs) in relation to their initialization and internal estimators. Utilizing a unique technique for the algebraic representation of a multiset of vectors, termed Multiset Equivalent Function (MEF), it is demonstrated that GNNs can generate functions satisfying a weight-equipped variant of permutation-equivariance using distinct features. The MEF technique further reveals that GNNs, equipped with node identifiers in their initialization, can generate any function in a fully connected and weighted graph scenario. These findings contribute to a formal understanding of the intricate relationship between GNNs and other algorithmic procedures applied to graphs, such as min-cut value and shortest path problems, as well as the re-derivation of the known connection between GNNs and the Weisfeiler-Lehman graph-isomorphism test.