Blockchain based Federated Deep Learning Framework for Malware Attacks Detection in IoT Devices
Over the recent years, billions of IoT devices that do not have adequate security features have been created and deployed over internet, and with the increased bandwidth and lower latency of 5G networks these numbers expected to grow exponentially. Consequently, there is a pressing requirement of reliable methods to identify malware infected IoT devices present in a network. Traditional Centralized Deep Learning approach was used to train a model capable of detecting infected IoT devices, but this approach faces major challenges like data privacy violation, scalability and high network latency. A more efficient process Federated Deep Learning was proposed which aims to improve privacy and security by keeping data on the devices and transmitting only model parameters to a central location for aggregation and then this aggregated model reflected to the edge devices. Because of the centralized aggregation this approach is prone to adversarial attacks. To resolve these vulnerabilities, we presented a decentralized framework consisting of a chain of Federated blocks at each edge devices with cohere-consensus mechanism (FBCC). The framework makes use of blockchain technology both for updating local models and to store global models. We also came up with a novel cohere-consensus technique to facilitate the suggested FBCC, which efficiently cut down on the quantity of consensus computing while simultaneously lowering the risk of malicious attacks. At last, we conducted comprehensive testing of the frameworks with real-world datasets, in both benign and malicious environments. The results revealed that our proposed framework outperformed the other frameworks in the malicious environment, while demonstrating comparable efficiency in the benign environment. Additionally, we conducted a thorough comparison of the frameworks in terms of memory requirements and training time. FBCC framework, due to its core architecture, exhibited slightly lower performance than FDL. However, it maintained privacy as a top priority.
Paper
Full text
Blockchain based Federated Deep Learning Framework for Malware Attacks Detection in IoT Devices
Semantic Scholar · Computer Science · 2023
Abstract
Over the recent years, billions of IoT devices that do not have adequate security features have been created and deployed over internet, and with the increased bandwidth and lower latency of 5G networks these numbers expected to grow exponentially. Consequently, there is a pressing requirement of reliable methods to identify malware infected IoT devices present in a network. Traditional Centralized Deep Learning approach was used to train a model capable of detecting infected IoT devices, but this approach faces major challenges like data privacy violation, scalability and high network latency. A more efficient process Federated Deep Learning was proposed which aims to improve privacy and security by keeping data on the devices and transmitting only model parameters to a central location for aggregation and then this aggregated model reflected to the edge devices. Because of the centralized aggregation this approach is prone to adversarial attacks. To resolve these vulnerabilities, we presented a decentralized framework consisting of a chain of Federated blocks at each edge devices with cohere-consensus mechanism (FBCC). The framework makes use of blockchain technology both for updating local models and to store global models. We also came up with a novel cohere-consensus technique to facilitate the suggested FBCC, which efficiently cut down on the quantity of consensus computing while simultaneously lowering the risk of malicious attacks. At last, we conducted comprehensive testing of the frameworks with real-world datasets, in both benign and malicious environments. The results revealed that our proposed framework outperformed the other frameworks in the malicious environment, while demonstrating comparable efficiency in the benign environment. Additionally, we conducted a thorough comparison of the frameworks in terms of memory requirements and training time. FBCC framework, due to its core architecture, exhibited slightly lower performance than FDL. However, it maintained privacy as a top priority.