Optimization of Privacy-Utility Trade-Off for Efficient Feature Selection of Secure Internet of Things

5G wireless network technology is widely used, and data protection is becoming more important as data transmitted and received over the network increases. As the number of Internet of Things devices rapidly increases, data leakage attacks targeting lightweight devices are increasing. Machine learning models have limitations in being applied to lightweight devices because of their large computational complexity and latency in the learning process. Improving utility while protecting data using existing data privacy protection techniques is difficult. Guaranteeing privacy and data utility is difficult because processing large amounts of data with low-capacity memory is complicated. Therefore, a memory-efficient mechanism is needed while ensuring privacy and data utility in data transmission conditions. We propose an optimal feature selection mechanism that maximizes privacy and utility by optimizing the privacy-utility trade-off for resource-constrained lightweight device environments. The proposed mechanism extracts privacy-sensitive features by selecting features requiring privacy protection, removing privacy-sensitive features, making it difficult for attackers to identify data even if they intercept them during data transmission. It demonstrated improved accuracy and memory usage compared to conventional models and improved the accuracy of legitimate nodes by 17.6% compared to models with differential privacy, reduced the accuracy of attackers by 15.45% compared to models without privacy protection techniques, and memory usage by 18.26% under 100% data sampling ratio conditions. Thus, it has been proven to be a secure and efficient model for data transmission environments by effectively improving data utility, privacy, and memory usage.