Friday 21 March 2025
In a recent study, researchers have developed a new type of attack that can compromise the security of decentralized federated learning systems. Decentralized federated learning is a technique used in machine learning where multiple devices or organizations work together to train a single model without sharing their raw data.
The new attack, known as DMPA (Decentralized Model Poisoning Attack), targets the collaborative nature of decentralized federated learning by manipulating the training process. The attackers create and disseminate compromised models that can affect the performance of the overall system.
To execute the attack, malicious clients create multiple models that are designed to mislead the training process. These models are then shared with other clients, who unknowingly incorporate them into their own models. As a result, the entire system becomes corrupted, compromising the accuracy and reliability of the trained model.
The researchers tested DMPA on three different datasets: CIFAR-10, MNIST, and Fashion-MNIST. They found that the attack was successful in all cases, with the compromised models achieving significantly lower F1 scores than the original models.
One of the most concerning aspects of DMPA is its ability to adapt to different network topologies. The researchers tested the attack on three different types of networks: fully connected, ring, and star. They found that the attack was successful in all cases, regardless of the network topology.
The study highlights the need for more robust security measures in decentralized federated learning systems. The researchers suggest that future work should focus on developing algorithms that can detect and mitigate model poisoning attacks like DMPA.
The findings also raise concerns about the potential impact of DMPA on real-world applications of decentralized federated learning. For example, in healthcare, compromised models could lead to inaccurate diagnoses or ineffective treatments.
Overall, the study demonstrates the importance of security considerations in the development of decentralized federated learning systems. As the technology continues to evolve, it is crucial that researchers and developers prioritize the protection of these systems against attacks like DMPA.
The study’s findings have significant implications for the future of machine learning and artificial intelligence. As more devices and organizations rely on decentralized federated learning, the need for robust security measures will only continue to grow.
Cite this article: “Decentralized Federated Learning Systems Vulnerable to Model Poisoning Attacks”, The Science Archive, 2025.
Decentralized Federated Learning, Machine Learning, Model Poisoning Attack, Artificial Intelligence, Cybersecurity, Data Security, Network Topology, F1 Scores, Compromised Models, Robust Security Measures
