PraaKrum: A Practical Byzantine-Resilient Federated Learning Algorithm
Prabhleen Kukreja, V. Mahendran · 2024
Federated Learning (FL) is considered as a suitable paradigm for intelligent data analytics over Internet of Thing (IoT) devices. While the data-privacy preserving feature of FL is useful, the lack of data auditing abilities of the participating IoTs creates a potential attack surface. To this end, a potential attacker can perform data/model poisoning attacks on the FL. To protect from such attacks, a popular theoretically proven algorithm for FL namely, Krum [1] was proposed. From an implementation perspective, the Krum algorithm needs to be provided with apriori information such as the total number of attackers in each round, which is impractical. It is not feasible to know, in advance, which of the participating IoT clients are attackers. Furthermore, in the presence of mobile Parameter Server (PS) which stays in the WiFi coverage area consisting of stationary clients for an unknown time, the implementation of Krum algorithm in such practical setting is not investigated, to the best of our knowledge. With an aim to provide a practical implementation of a resilient FL framework, for the first time, this paper investigates the sensitivity of the attackers information towards the robustness offered by the Krum-based FL algorithm. The insights gained from the sensitivity analysis are used to develop a novel and Practical algorithm of Krum (PraaKrum) for FL for a WiFi network with mobile PS. The developed analysis is helpful in both understanding the key resilient characteristics offered by the Krum algorithm and providing the research community with a guideline towards the practical implementation of the attacker-resilient Krum-based FL algorithm in a mobile wireless setting.