The Cost of Privacy: A Comprehensive Analysis of the Security Issues in Federated Learning

Federated Learning (FL) is an emerging subclass of Artificial Intelligence that decentralizes the learning process. It allows participants access to an extensive model trained in a broader dataset without ever sharing their private data directly. FL combines multiple local models into a global model, thereby diminishing the need for individual participants to have large datasets. This decentralized nature of FL makes it more susceptible to adversarial attacks. Adversaries generally try to infect a portion of the FL model, altering the classification of some categories, known as a backdoor attack, or can try to completely render the global model useless by reducing the overall model accuracy, known as a byzantine attack. Detecting and defending against byzantine and backdoor attacks in an FL system is difficult since the adversary can also act as a participant trying to disrupt the system. These security concerns have made adopting FL models complex on a global scale. Many researchers have exploited these security concerns in the past few years and developed numerous attacks based on multiple vulnerabilities. In this chapter, we comprehensively analyze the diverse elements that make each security attack different from another, and look at the native FL characteristics that make it vulnerable to such attacks. We will also discuss the defense tactics that can be used to identify malicious participants or mitigate the impact of an attack in the global model.