Intrusion detection using synaptic intelligent convolutional neural networks for dynamic Internet of Things environments

The swift proliferation of IoT devices has brought about a multitude of complex cyberattacks that breach network security and compromise user privacy. To address these threats, this paper proposes a synaptic intelligent convolutional neural network (SICNN) model for intrusion detection in dynamic IoT environments. Confronted with real-time changing intrusion data, numerous intrusion detection systems necessitate the continuous integration of new training data for model retraining and parameter refinement. Nevertheless, the storage demands brought about by continuous input data streams and the time consumed by repetitive training pose significant challenges for IoT intrusion detection. The SICNN model leverages the synaptic intelligence (SI) algorithm to optimize the synaptic structure of the convolutional neural network (CNN), significantly mitigating the forgetfulness of the CNN for past detection tasks and simplifying model training. Additionally, a novel loss function is designed to address the class imbalance present in IoT traffic data and to mitigate the gradient vanishing issue observed in traditional loss functions. Enhancing the deployability of the detection model on resource-limited IoT devices, the model's structure and parameters are quantized, and the loss associated with model quantization is minimized through quantization-aware training. Ultimately, experiments are conducted on the intrusion detection datasets CIC IDS2017 and CICIoT2023, demonstrating that the proposed detection method outperforms other state-of-the-art intrusion detection approaches. These results underscore the relevance and effectiveness of the SICNN model in enhancing IoT security, making it a promising solution for real-time intrusion detection in dynamic IoT environments.