Adaptive and Secure Coded Computation for Distributed Learning

The volume of data generated by end devices such as Internet of Things (IoT) has been growing exponentially. To unlock the potential of this vast data, particularly in machine learning applications, distributed edge computing has emerged as a promising paradigm. In this approach, computationally intensive tasks are distributed across edge devices, with a central server aggregating the results. However, traditional distributed computing methods often yield suboptimal performance due to resource constraints at edge devices, the presence of straggler devices, and the need to protect sensitive data. To address these challenges in edge networks, this thesis designs and develops distributed edge computing solutions using tools from coded computing, Multi-Party Computation (MPC), and information theoretic security. First, we focus on characterizing the cost-benefit trade-offs of coded computation in practical edge computing systems and develop a resource aware adaptive coding strategy for matrix multiplication by taking into account the heterogeneous and time varying nature of edge devices. Next, we investigate efficient coded multi-party computation at edge networks. We designed novel CMPC algorithms PolyDot-CMPC and AGE-CMPC to minimize the number of edge devices, storage, communication, and computation load as compared to existing CMPC mechanisms. Finally, we focus on privacy preserving vertical federated learning. We designed a set of privacy preserving tools including low-complexity dot-product calculation and binary version of Shamir's secret sharing to provide an efficient and practical solution for collaborative machine learning in vertical federated learning setup while ensuring data privacy.