Newly published paper presents breakthrough in communication-efficient and robust Peer-to-Peer Federated Learning

Researchers from the ARCADIAN-IoT project partner RISE have recently published a groundbreaking paper in Computers & Security, Volume 129, June 2023, titled “SparSFA: Towards robust and communication-efficient peer-to-peer federated learning” The paper presents a novel communication-efficient and robust weighting scheme called SparSFA that mitigates the impacts of adversarial attacks in a peer-to-peer federated learning (P2PFL) setup.

Federated learning (FL) is an innovative approach to training collaborative machine learning (ML) models while preserving the privacy of participants’ datasets. Traditional FL methods, however, have limitations that can hinder their applicability, especially in scenarios with limited connectivity, such as IoT applications. In such cases, the need for a server or aggregator to orchestrate the learning process may not be feasible, and these methods offer less flexibility for personalizing ML models for different participants.

Peer-to-peer federated learning (P2PFL) addresses these limitations by providing more flexibility, allowing participants to train their own models in collaboration with their neighbors. However, the communication burden of this approach can be significant due to the large number of parameters in typical deep neural network architectures. Furthermore, standard aggregation schemes for FL are highly susceptible to data and model poisoning attacks.

To combat these issues, the RISE researchers developed SparSFA, an algorithm for P2PFL capable of reducing communication costs. SparSFA demonstrates outstanding performance among four state-of-the-art aggregation methods in four different attack scenarios, including data poisoning and model poisoning attacks. It also effectively addresses anomaly detection problems in any random network topology.

The researchers’ empirical evaluation on real datasets for intrusion detection in IoT, considering both balanced and imbalanced-dataset scenarios, shows that SparSFA is highly robust against various indiscriminate poisoning attacks launched by one or multiple adversaries. Furthermore, it outperforms other robust aggregation methods while reducing communication costs through sparsification.

The SparSFA algorithm has the potential to significantly impact the future of federated learning, particularly in IoT applications, by improving communication efficiency and providing robustness against adversarial attacks. This breakthrough has far-reaching implications for industries and applications that rely on secure and efficient data sharing and collaborative machine learning.

Share this