Work

SecFedDrive: Securing Federated Learning for Autonomous Driving Against Backdoor Attacks

Year: 2024

Type: conference-paper

Abstract: Federated learning (FL) enables collaborative model training without sharing the private data of each individual participant, making it well-suited for autonomous driving applications. Preserving the integrity of sensitive data is crucial for the security of these systems due to the direct implications for passenger safety. Although federated model training enhances data privacy for individual participants, it remains vulnerable to stealthy backdoor attacks that can alter the global model, potentially threatening system reliability in high-stakes scenarios such as autonomous driving. We introduce SecFedDrive, a robust defense mechanism against backdoor attacks on a decentralized deep FL system. Our approach reduces attack success rates to as low as 2%, significantly enhancing the security and reliability of FL models in autonomous driving environments, and demonstrates significant reductions in training time relative to a comparable architecture.

Source: IEEE Conference on Communications and Network Security

DOI: 10.1109/cns62487.2024.10735501

Authors: Rahul Kumar, Gabrielle Ebbrecht, Junaid Farooq, Wenqi Wei, Ying Mao, Juntao Chen

Institution: University of Michigan–Dearborn

Language: English

Topic: Adversarial Robustness in Machine Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Funders: National Science Foundation, Fordham University, Division of Electrical, Communications and Cyber Systems

Award: ERI: Multi-Layer Dynamic Strategic Decision-Making for Integrated Cyber-Physical Energy Systems Security and Resilience