FLAMINGO: Adaptive and Resilient Federated Meta-Learning against Adversarial Attacks

Md. Zarif Hossain, Ahmed Imteaj, Abdur Rahman Bin Shahid

Abstract: In today’s data-centric world, the synergy between Meta Learning and Federated Learning (FL) signifies a new era of technological advancement, driving rapid adaptation, improved model generalization, and collaborative model training across decentralized networks. This fusion, known as Federated Meta-Learning (FML), emerges as a cutting-edge solution for resource-constrained edge devices, enabling the production of personalized models with limited training data. However, FML navigates a complex terrain, balancing efficiency with security, as adversarial attacks on edge devices pose significant threats. These attacks risk introducing bias and undermining the integrity of model training, a critical concern given the typically sparse data on edge devices. This paper explores the intricate dynamics of FML amidst such adversarial challenges, introducing a novel algorithm, FLAMINGO. FLAMINGO is designed to conduct adversarial meta-training coupling with data augmentation and consistency regularization strategies, thereby strengthening the meta-learner’s defenses against adversarial attacks. This strategic approach not only protects meta-learners against adversarial threats but also prevents overfitting, striving a balance between privacy, security, and technological efficiency, all while optimizing communication costs in the FML landscape. We have released our code on GitHub1, which is publicly accessible.1https://github.com/speedlab-git/Flamingo-Adversarial-FML.git

External IDs:dblp:conf/icdcs/HossainIS24