Technical Challenges and System Limitations
Abstract
Federated learning (FL), which is the main model of decentralized collaborative learning (DCL), has become a ground breaking way to solve the fundamental problem that arises when you combine the data-heavy nature of modern artificial intelligence (AI) with strict legal and moral requirements for data privacy and security. According to DCL, they can achieve collaborative intelligence without storing sensitive raw data in one place by training models locally on distributed devices and only collecting model updates. This chapter examines this paradigm critically and argues that, although it effectively addresses the issue of data privacy in transit, it is not without its fair share of challenging technical issues and limitations at the system level. Here are the three primary areas into which we divide these issues: (1) non-identically distributed (non-IID) data across clients causes model divergence and instability; (2) communication overhead causes a big bottleneck in environments with limited resources; (3) security and privacy vulnerabilities make the decentralized architecture more vulnerable to model poisoning, Byzantine failures, and sophisticated inference attacks that can re-create sensitive training data using shared model updates. Some mathematical ways to think about these problems in this chapter include the Federated Averaging algorithm, metrics for statistical divergence, and models for attack vectors. We illustrate the performance drop due to these problems and the costs and benefits of fixing them with descriptive graphs and diagrams. Specific examples from healthcare, mobile computing, and industrial IoT show how these technological challenges manifest in practice. We draw the conclusion that DCL is more of a complicated system of trade-offs than a “plug-and-play” confidentiality solution. To build a distributed learning ecosystem that is secure, efficient, and trustworthy, it is necessary to take a comprehensive approach that co-designs algorithmic solutions (such as robust aggregation rules), communication efficiencies (such as model quantization), and cryptographic safeguards (such as differential privacy and secure aggregation).