Detection and Mitigation Data Poisoning Attacks in Multimodal Online Federated Learning

The Internet of Things (IoT) ecosystem produces vast quantities of multimodal data from diverse sources such as sensors, cameras, and microphones. With the growing integration of edge intelligence, IoT devices have evolved beyond simple data collection units into intelligent edge nodes, necessitating distributed learning paradigms to handle heterogeneous and dynamic multimodal data effectively. Moreover, the real-time nature of data generation and the limited storage capabilities of edge devices call for an online learning framework. To address these challenges, Multimodal Online Federated Learning (MMO-FL) has emerged as a promising solution, enabling decentralized and real-time model training across multiple modalities. However, existing MMO-FL studies largely assume idealized environments and overlook security threats. As a result, the security landscape of MMO-FL remains severely underexplored. In practice, MMO-FL introduces unique and complex security vulnerabilities across three dimensions: spatial (federated), temporal (online), and modal (multimodal). These characteristics make the distributed and online data collection process highly vulnerable to adversarial manipulation. To bridge this gap, we present the first systematic study of data poisoning attacks within the MMO-FL scenario. We begin with a theoretical analysis quantifying the impact of such attacks on learning performance. To defend against them, we propose a novel detection and mitigation algorithm tailored specifically for MMO-FL systems. Extensive experiments conducted on two real-world multimodal datasets, UCI-HAR and USC-HAD, demonstrate that our approach effectively detects and mitigates data poisoning attacks.