XNetIoT: An Extreme Quantized Neural Network Architecture for IoT Environment Using P4
Suvrima Datta, Aditya Kotha, U. Venkanna, K. Mallikharjuna Rao · IEEE Transactions on Network and Service Management · 2024
Internet of Things (IoT) security has heavily relied on Machine Learning (ML) techniques in recent years. However, these techniques inherently require a vast amount of labeled data for training, which is challenging. Moreover, these techniques are also prone to data shift problems. A practical approach to address these challenges can be through using self-learning. Hence, this paper proposes XNetIoT, an extremely quantized neural network architecture, to enhance the security of IoT networks sequentially using a Data Plane (DP). The proposed solution is two-fold: primarily, XNetIoT focuses on the binary in-network classification of incoming IoT traffic flows in the DP. Additionally, it accumulates the real-time features for updating the XNetIoT and is sent to the control plane. Within the control plane, the XNetIoT is retrained with accumulated real-time features. Further, the control plane is used to identify the specific attack types in the attack traffic. Subsequently, appropriate action is installed in the DP to mitigate the attacks. The comprehensive evaluation of our proposed solution yields 98.44% and 99.22% accuracy for in-network binary and multiclass classification using XNetIoT. Moreover, to assess the adaptability of XNetIoT, our model was trained using several distinct types of attacks and was subsequently evaluated through testing with unseen attacks, which are not part of training or testing. Interestingly, XNetIoT achieved an accuracy of 98.47% for detecting unknown attacks. Further, the average attack traffic flow classification time is 0.2475ms in XNetIoT.