A multi-level taxonomy-driven survey of security threats and defenses in federated learning across silos and devices

Federated learning (FL) provides efficient and secure training by retaining data at its sources and sharing only model updates, thereby enhancing privacy and reducing security risks. However, FL is vulnerable to a wide range of attacks at multiple layers, and security techniques have been integrated to ensure a secure end-to-end FL process. This survey goes beyond existing works by providing a unified, end-to-end security perspective for FL, introducing a security-oriented, multi-level taxonomy-driven survey that jointly considers FL layers, threats, adversarial models, and deployment scenarios. We analyze the threat models, security requirements, and challenges at each layer, including the client, communication, and server/aggregation layers. We provide a comprehensive review of state-of-the-art security techniques employed throughout the FL process. We present a layer-based attack and its corresponding defense mechanisms tailored to cross-silo and cross-device settings. Most notably, this survey provides a systematic classification and joint mapping of security solutions to the diverse layer-specific security requirements of FL systems, highlighting trade-offs and suitability across different deployment scenarios. Additionally, we review existing open-source FL frameworks with integrated security features and identify key limitations, open challenges, and promising future research directions to guide the development of robust, scalable, and trustworthy FL systems.