A lightweight integrity-driven federated learning approach to mitigate poisoning attacks in IoT

Latif, Shahid; Djenouri, Djamel; Hernandez-Ramos, Jose L; Skarmeta, Antonio; Ahmad, Jawad

Abstract

Despite its distributed nature and being privacy-preserving by nature, Federated Learning (FL) is vulnerable to poisoning attacks in which malicious actors can inject fake model parameters or false data to compromise the learning process. This article introduces a lightweight and efficient integrity verification scheme to mitigate these attacks on FL platforms in Internet of Things (IoT) networks. The core design of the proposed scheme is based on a customized feed-forward neural network (FFNN) integrated with a fast, secure, and efficient Keccak-512 hashing algorithm. This combination balances security, speed, efficiency, and suitability for resource-constrained IoT devices. The proposed model was trained and evaluated using the real-time IDSIoT2024 dataset, and the results demonstrated a higher classification accuracy of 98.29% with a lower memory footprint of 87.58KB. Furthermore, the lower computation and communication overheads, low CPU and GPU memory utilization confirm the resource and time efficiency of the proposed scheme to effectively mitigate the poisoning attacks in FL architectures.

• Presentation Conference Type: Conference Paper (unpublished)
• Conference Name: 2024 IEEE Future Networks World Forum (FNWF)
• Start Date: Oct 15, 2024
• End Date: Oct 17, 2024
• Acceptance Date: Sep 15, 2024
• Deposit Date: Sep 18, 2024
• Peer Reviewed: Peer Reviewed
• Keywords: Cybersecurity; Federated Learning; Integrity Verification; IoT; Poisoning Attacks
• Public URL: https://uwe-repository.worktribe.com/output/12895988
• Additional Information: This work is part of the REMINDER project, funded under the EU CHIST-ERA initiative (Grant EP/Y036301/1 from EPSRC, UK). It is also partially supported by the HORIZON-MSCA-2021-PF-01-01 project, INCENTIVE (Grant Agreement 101065524), and a 2023 Leonardo Grant from the BBVA Foundation.