Intrusion detection system based on federated learning using layer-based privacy obfuscation

Intrusion detection system based on federated learning using layer-based privacy obfuscation

연구 분야: Analysis

학회: Cluster Computing

In light of a rapid evolution of Industrial Internet of Things (IIoT), the issues of securing these integrated systems with dealing scalability and privacy becoming more crucial. To minimize the privacy attacks in Federated Learning (FedLrn) based Intrusion Detection Systems (IDSs) for IIoT and to optimize the negative influence on model that employed standard privacy preserving mechanisms to prevent privacy attacks. This study deploys a novel approach that IDS based on FedLrn with Layer-Based Privacy Obfuscation Mechanism (LB-POM), intended to improve privacy, utility and efficiency in IIoT setting. FedLrnLB-POM approach includes FedLrn training process across IIoT devices without sharing sensitive information. The significant contribution of this approach is the utilization of LB-POM, that specifically preserve the privacy of significant model layers (sensitive layers) which dedicatedly detect intrusions. For privacy preservation of sensitive layers, LB-POM employs Adaptive Homomorphic Encryption and Predictive Error Minimization (AHEnc-PrEMin), while non sensitive layers’ s parameters are communicated transparently. This approach ensures that only sensitive information is communicated with preserved privacy, while non sensitive information communicate in plaintext that results to improving privacy along with reducing overhead. In addition to this, it is addressed that FedLrn aggregation methods in existing research facing issues of privacy threats, communication bottlenecks, limiting their scalability, therefore authors employ a hybrid aggregation mechanism that Federated Hybrid Quintile Mixture Aggregation (FedHQMA) adapts its quantile selection based on the statistical properties of client updates, specifically skewness and variance. This ensures that FedHQMA is more robust to non-IID client distributions and Byzantine failures. With CICToN_IoT, UNSW15 and Bot_IoT dataset, authors perform an experiment and evaluate results. The results demonstrate that privacy, accuracy and computational time of the proposed FedLrnLB-POM attained accuracy is (At = 95.5 percent) at privacy budget (ϵ = 0.9) with training time (t = 120 ms). FedHQMA shows accuracy 95 percent, while communication efficiency 91 percent and computational efficiency 90 percent. With respect to AHEnc-PrEMin, the privacy and accuracy trade off shows that the Biweight Midcorrelation Coefficient (BMCC) is 0.92 at privacy budget (ϵ = 0.9). On the basis of results, it is analyzed that FedLrnLB-POM outperforms the other state of art methods.