This paper investigates the impact of locally differentially private mechanisms on causal discovery algorithms. The authors consider two main approaches: local differential privacy (LDP) represented by the k-Ary Randomized Response (k-RR) mechanism, and local d-privacy represented by the Geometric mechanism.
The authors conduct extensive experiments on both real and synthetic data sets, evaluating the performance of 9 causal discovery algorithms, including constraint-based, score-based, and causal asymmetry-based methods. The key findings are:
Locally d-private mechanisms, such as the Geometric mechanism, generally outperform LDP mechanisms like k-RR in preserving the causal structure of multidimensional data. The Geometric mechanism maintains the performance of causal discovery algorithms close to the non-privatized data, while k-RR noise deteriorates the data structure and leads to worse algorithm performance.
The authors introduce a unified privacy measure from an attacking perspective, allowing for the comparison of LDP and local d-privacy. This measure facilitates the assessment of privacy-utility trade-offs in real-world tasks such as causal discovery.
For smaller multidimensional data sets, the variation in performance is high, making it difficult to draw reliable conclusions. However, the authors still observe a slight advantage in applying Geometric mechanisms over k-RR on the Synth5 and Human Stature data sets.
On the two-dimensional CEP data set, the Geometric mechanism consistently outperforms k-RR, with notable improvements, especially in the case of the RECI algorithm, where the accuracy surpasses the baseline.
The authors conclude that their findings provide valuable insights into the application of locally private mechanisms in real-world causal discovery tasks, aiding practitioners in collecting multidimensional user data in a privacy-preserving manner.
To Another Language
from source content
arxiv.org
Deeper Inquiries