Local Differential Privacy Preservation via the Novel Encoding Method

In recent years, the application of big data analysis has seen a significant increase across various fields with the aim of extracting useful information from massive data to enhance human life. These data contain a large amount of privacy, and directly using user data for analysis poses a significant risk of privacy leakage. To address this pressing issue, many scholars have incorporated the latest protection methodologies, such as local differential privacy (LDP), into big data analysis. Existing LDP methods typically convert data into binary strings and then add noise, but the weight of each bit unevenly distributes noise, eventually increasing utility loss. Inspired by this, the paper investigates the reduction of error through alternative numeral systems, leading to the proposition of a segmented-based LDP data preservation mechanism (LDPseg), where each bit flip has an equal impact on the outcome. Theoretical analysis reveals that under certain conditions, this mechanism maintains a lower error expectation and variance compared to binary systems. Real-world experimental results indicate that the proposed method exhibits positive performance in machine learning.