Comparison of LSTM and MLP Trained Under Differential Privacy for Intrusion Detection

There is an escalation in cyber-attacks on cyber-physical systems and IoT devices. Intrusion detection systems are becoming increasingly crucial in detecting and thwarting adversarial attacks. Machine learning algorithms, including deep learning, are used for intrusion detection. However, datasets used for training machine learning models for intrusion detection may contain information sensitive to an organization. Privacy attacks such as membership inference, property inference, and reconstruction attacks to an intrusion detection system can cause the disclosure of sensitive information about the organization's network-preserving machine learning techniques, which have been used to address privacy threats. In this paper, we train the deep learning models:-Long-term memory (LSTM) and multi-layer perceptron (MLP) with differential privacy for intrusion detection. We use a differentially private stochastic gradient descent (DP-SGD) algorithm in model training. We conduct experiments on four different IoT network datasets and compare the two models on their classification accuracy and privacy budget. Our experiments show that the LSTM trained under differential privacy had better classification accuracy, precision, and f1 score than the MLP trained under differential privacy. Conversely, the MLP had a tighter privacy budget given the same noise level. This research provides insights into selecting a privacy-preserving deep learning model for intrusion detection.