Overview
RoadFed is a multimodal federated learning system designed for the detection and early alarm of road hazards within Collaborative Intelligent Transportation Systems (C-ITS). The system integrates a multimodal road hazard detector, a communication-efficient federated learning approach, and a local differential privacy method tailored for high-dimensional multimodal data.
Research Context
Internet of Things (IoTs) have found broad application in C-ITS for supporting road accident prevention. A significant factor in road accidents within C-ITS is the need for efficient detection and early warning of road hazards. While existing research has explored this area and yielded results, many solutions primarily utilize single-modality data. Additionally, current approaches often encounter challenges such as high computation and communication overhead, or issues related to the curse of high dimensionality in their privacy-preserving methodologies. RoadFed was developed to address these specific obstacles.
Approach
The RoadFed framework was designed with three core components:
- Innovative Multimodal Road Hazard Detector: This component is central to the system's ability to process and interpret data from multiple modalities.
- Communication-efficient Federated Learning Approach: This element focuses on optimizing the communication overhead inherent in collaborative learning processes.
- Customized Low-error-rate Local Differential Privacy Method: This method was specifically crafted to manage high-dimensional multimodal data while ensuring privacy preservation.
The system is engineered to facilitate collaborative training across various data modalities on multiple edge devices, utilizing non-iid (non-independent and identically distributed) high-dimensional real-world data, all while maintaining privacy preservation for road users.
Findings
Experimental evaluations of RoadFed were conducted using both self-gathered real-world datasets and the public CrisisMMD dataset. The results indicated that RoadFed outperformed most existing systems in this domain. Key findings include:
- RoadFed achieved an accuracy of 96.42% in road hazard detection.
- The system demonstrated a latency of 0.0351 seconds.
- Communication cost was observed to be up to 1,000 times lower compared to existing systems in the field.
These results collectively suggest that RoadFed effectively addresses challenges related to single-modality data processing, computational and communication overhead, and privacy concerns associated with high-dimensional data within C-ITS.