SUMO Simulations for Federated Learning in Communicating Autonomous Vehicles
summary
Uses the Monaco SUMO Traffic Scenario (MoST) to record parking occupancy with 60 simulation repeats producing about ~0.25 TB data, 60 runs, trains identical neural nets (MSE loss, RMSprop, early stopping) for centralized, individual and federated schemes, and measures privacy by trying to infer which of five parking lots a vehicle visits and its timing. Centralized achieves best accuracy with low perceived privacy risk; individual leaks location/time (≈8 minute timing error, can identify 5/5 visited lots); federated improves accuracy over individual but initially reveals on average ~2.5 of 5 known lots and degrades with more participants. Also describes a Dockerized SUMO client-server setup to distribute simulations and GPU training.
AI-written summary · about
chapters (9)
- 0:00Data collection and ML
- 1:53Privacy and security risks
- 3:34Centralized vs individual
- 5:10Traffic simulation setup
- 6:54Technical model execution
- 9:28Individual model performance
- 11:57Tracking vulnerabilities
- 13:45Federated learning analysis
- 16:59Conclusion and outlook
related talks
13mExtending SUMO for Lane-Free Microscopic Simulation of Connected and Automated Vehicles
25mDriver Models in SUMO and Elsewhere
20mSimulating platooned connected autonomous vehicle in a large scale urban scenario
34mDecentralized Machine Learning
41mFederated Learning: Machine Learning on Decentralized Data
45mKEYNOTE: Traffic simulation modeling: combining generative modeling with transportation science ...
27mFederated Learning with Azure Machine Learning NVIDIA FLARE and MONAI
19mTestbed for Development of AI Solutions to Boost Operational Efficiency in Transportation Networks