Threat Landscape in Federated Learning

Although artificial intelligence (AI)-enabled Internet-of-Things (IoT) systems increasingly employ ML to provide localized intelligence, the rising heterogeneity and dynamic characteristics of IoT data highlight the limitations of task-specific models. FL with the ability to generalize across domains and tasks through large-scale pretraining is being widely adopted to support IoT applications, such as predictive maintenance, healthcare monitoring, and autonomous mobility. Moreover, FL facilitates cross-domain knowledge transfer, thereby reducing the need to train models from scratch on every device. However, integrating FL into IoT introduces unique security concerns: unlike conventional AI systems, FL operates in distributed and resource-constrained environments, often relying on federated or decentralized training across IoT nodes. This distributed interaction exposes FL to IoT-specific threats, including model poisoning (MP), inference attacks, and adversarial manipulations, particularly under non-independent and identically distributed (non-IID) data distributions and intermittent connectivity. Consequently, addressing the security challenges of FL within IoT ecosystems is crucial for enabling trustworthy AI and ensuring the resilience of IoT infrastructures.