Machine learning (ML) techniques have brought revolutionary impact to human society, and they will continue to act as technological innovators in the future. To broaden its impact, it is urgent to solve the emerging and critical challenges in machine learning, such as efficiency and security issues. On the one hand, ML models have become increasingly powerful due to big data and models, but it also brings tremendous challenges in designing efficient optimization algorithms to train the big ML models from big data. The most effective way for large-scale ML is to parallelize the computation tasks on distributed systems composed of many computational devices. However, in practice, the scalability and efficiency of the systems are greatly limited by information synchronization since the message passing between the devices dominates the total running time. In other words, the major bottleneck lies in the high communication cost between devices, especially when the scale of the system and the models becomes larger while the communication bandwidth is relatively limited. This communication bottleneck often limits the practical speedup of distributed ML systems. On the other hand, recent research has generally revealed that many ML models suffer from security vulnerabilities. In particular, deep learning models can be easily deceived by the unnoticeable perturbations in data. Meanwhile, graph is a kind of prevalent data structure for many real-world data that encodes pairwise relations between entities such as social networks, transportation networks, and chemical molecules. Graph neural networks (GNNs) generalize and extend the representation learning power of traditional deep neural networks (DNNs) from regular grids, such as image, video, and text, to irregular graph-structured data through message passing frameworks. Therefore, many important applications on these data can be treated as computational tasks on graphs, such as recommender systems, social network analysis, traffic prediction, etc. Unfortunately, the vulnerability of deep learning models also translates to GNNs, which raises significant concerns about their applications, especially in safety-critical areas. Therefore, it is critical to design intrinsically secure ML models for graph-structured data.The primary objective of this dissertation is to figure out the solutions to solve these challenges via innovative research and principled methods. In particular, we propose multiple distributed optimization algorithms with efficient message passing to mitigate the communication bottleneck and speed up ML model training in distributed ML systems. We also propose multiple secure message passing schemes as the building blocks of graph neural networks aiming to significantly improve the security and robustness of ML models.
Read
