Enhancing IoT Security: Federated Learning with Autoencoder Model for IoT Attacks Detection
Rabiya Majeed, Amrit Lal Sangal · 2024
Now a days, Internet of Things(IoT) devices are mass produced because of their wide variety of applications in our daily lives, be it smart cities, smart homes, healthcare, IoT devices are used everywhere. IoT devices are heterogeneous in nature i.e, they vary in terms of underlying protocols and operating systems. This collection of devices produces a lot of data containing user information, which eventually exposes the entire system to attacks. IoT devices can remain out of sight for long periods of time, but like every other device they require regular updates to cope up with the evolving technologies, daily challenges and objectives. Although there are many concerns with the IoT device systems, the two main concerns are: data privacy and security. The strategy for protecting privacy explained in this paper focuses on making sure that no data is moved or transferred from the device off the network edge, instead the model is trained locally. The security issues are addressed by using a FL-based approach that employs an auto-encoder to detect botnet attacks using the decentralized traffic data. In order to detect botnet attacks, we have suggested a Federated Learning(FL)-based method that uses the network data from 9 different IoT devices and employs a deep auto-encoder model for training and testing. We demonstrate that we can obtain a 99% or greater accuracy in detection of malicious traffic using our proposed model. The entire comparative performance analysis shows a notable increase in attack detection accuracy rate between our decentralised suggested technique and a centralised structure.