Exploiting attribute correlation for reconstruction attacks on differentially private multi-attributed data

Differential Privacy (DP) is a widely used data privacy-preserving technique with single-attribute DP being a common approach, in which manipulated noise is applied to each data attribute individually. However, data in practical scenarios often contains multiple data attributes, and the correlations between these attributes, which are often overlooked, introduce vulnerabilities to single-attribute DP schemes. In this paper, we present a rigorous analysis demonstrating that these correlations can undermine the protection offered by single-attribute DP schemes, with the risk of compromise increasing as the correlation between attributes becomes more pronounced. We propose a novel attack framework to evade the single-attribute DP protection on multi-attributed data by exploiting the overlooked data attribute correlations. We further implement the attack by developing Machine Learning (ML) algorithms to uncover the straightforward and hidden attribute correlations. Extensive experiments with various ML algorithms are conducted to corroborate our analysis, demonstrating the existence of privacy leakage caused by data attribute correlations and the effectiveness of the proposed attack with significantly enhanced reconstruction accuracy. In one of our experiments, the proposed attack method mitigated over 50% of the DP noise, significantly enhancing the accuracy of reconstruction attacks.