A Federated Learning Monitoring Tool for Self-Driving Car Simulation (Student Abstract)

Taejoon Lee, Hyunsu Mun, Youngseok LeeAAAI 2023 · DOI

Abstract

Federated Learning (FL) offers a privacy-preserving framework for training autonomous driving models, yet debugging distributed simulations remains a significant challenge. This paper introduces CARLA-FLMon, a monitoring tool that integrates the CARLA simulator with the Flower FL framework. By providing real-time visualization of training metrics and client-side data distribution, CARLA-FLMon enables researchers to identify and exclude "low-contributing" clients. Preliminary tests on lane recognition demonstrate that removing a single client with invalid data can improve global model accuracy from 80.33% to 93.82%.

Introduction

Federated Learning (FL) is a promising approach for training autonomous driving models because it preserves privacy: vehicles train models locally on their own sensor data and only share model updates (weights) with a central server. However, FL is notoriously difficult to debug. In a distributed simulation, it is hard to determine if a drop in global model accuracy is due to poor hyperparameters, specific clients having "noisy" or invalid data, or hardware resource constraints.

Existing FL visualization tools are often designed for general-purpose machine learning and lack deep integration with the environmental variables of a driving simulator, such as weather, traffic, or sensor types. CARLA-FLMon addresses this gap by creating a tool specifically tailored to the CARLA ecosystem, allowing researchers to observe exactly what each simulated vehicle sees and how that data affects the global model.

System Architecture

CARLA-FLMon integrates the CARLA simulator with Flower, an open-source federated learning framework. The architecture consists of several key modules designed to streamline the experimental workflow:

  • FL Control Modules: These allow researchers to set simulation variables (e.g., weather, maps) via Python APIs and configure FL strategies (like FedAvg) and model parameters.
  • FL Monitor Modules: A web-based dashboard displays real-time metrics, including server and client-specific accuracy, loss, and epoch progress.
  • Client Data Monitoring: This module allows users to inspect the actual training data (images) being used by individual clients to identify anomalies.
The CARLA-FLMon system architecture and integrated dashboard interface.

Experimental Evaluation

The utility of CARLA-FLMon was tested using a lane recognition task within a simulated city map under sunny conditions. Data was collected from three simulated "clients" (cars). The study compared a baseline FL run using all three clients against an optimized run where a problematic client, identified via the monitoring tool, was excluded.

Results and Discussion

The initial experimental run with three clients achieved a global model accuracy of 80.33%. Using the CARLA-FLMon dashboard, researchers observed that Client 1 exhibited significantly lower accuracy than the others. By utilizing the tool's data inspection feature, the team discovered that Client 1 was processing a high volume of invalid or poor-quality images, which was "poisoning" the global model.

CARLA-FLMon's data inspection interface identifying invalid data from Car #1 for exclusion.

After excluding Client 1 and retraining with only the two "healthy" clients, the global model accuracy increased to 93.82%. This demonstrates that in FL, "less is more"—removing a bad actor can result in a much stronger global model than simply including more data. The tool successfully visualized resource utilization and learning progress across 20 global rounds in real-time, reducing the friction of iterative experimentation.

Conclusion and Future Work

CARLA-FLMon provides a vital bridge between autonomous driving simulators and federated learning frameworks. Future work will focus on expanding features to include more robust resource monitoring (CPU/GPU/Memory) and extending the tool to support more complex tasks like object recognition and safe driving scenarios involving traffic management. While the current prototype proved effective for small-scale simulations, its scalability to hundreds of concurrent clients remains a subject for future research.

Summary updated

BibTeX
@inproceedings{lee2023federated,
  author = {Taejoon Lee and Hyunsu Mun and Youngseok Lee},
  title = {{A Federated Learning Monitoring Tool for Self-Driving Car Simulation (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26984},
}
Figure