Analyzing Federated Learning From a Security Perspective
Akarsh K. Nair, Jayakrushna Sahoo, Ebin Deni Raj · Apple Academic Press eBooks · 2024
In the current technological scenario, privacy and security-related issues have been a significant concern for most companies working with vast volumes of data with high monetary value. User expectations associated with privacy have also increased drastically, forcing companies and service providers to shift to methodologies that ensure enhanced user and data privacy. Based on these issues, traditional artificial intelligence (AI) has also been transitioning from centralized to decentralized approaches as part of measures to preserve privacy. One such approach is federated learning, a type of distributed learning paradigm that aims to improve privacy by localizing data storage and conducting on-site training, thereby minimizing data transfer and reducing the risk of privacy breaches. However, even federated learning is not foolproof and is susceptible to security breaches. Adversarial attacks, such as poisoning attacks and leakage attacks, can compromise the performance of the model. Moreover, advancements in AI-based technologies have shown that data leakages in such systems can lead to the reconstruction of private data and models. This article provides a comprehensive analysis of various privacy issues associated with federated learning and presents proven methodologies that serve as preventive measures in adversarial scenarios. We discuss the effectiveness of attacks such as poisoning attacks, reconstruction attacks, backdoor attacks, and GAN-based attacks in federated learning systems. 214 Additionally, we explore future research directions to further enhance the security of this technology.