Differentially Private Top-$k$ Flows Estimation Mechanism in Network Traffic.

In network management, top- $k$ flows estimation of data streams is a fundamental task which has been extensively employed in traffic engineering, anomaly detection, and congestion control. To improve network quality, network measurement data need to be transferred between different network entities. However, it is challenging how to prevent the private information leakage when sharing data. In this paper, we put forward a local differential privacy (LDP) mechanism for finding top- $k$ flows among multiple independent clients. As most flows belong to a few flow identifiers, the flows of each client can be represented as a sparse vector. In our proposed scheme, we first present a high-utility LDP traffic aggregation scheme based on Hyperloglog sketch to accommodate the sparsity property of network flow data, and then utilize an approximate method of multi-round iterations to cut down the computation cost. We formally prove the proposed mechanism satisfies $(\epsilon, \delta)$ -LDP for $\iota$ -neighboring and compute its total error bound. Additionally, we evaluate our scheme by extensive experiments on both real-world and synthetic datasets, which indicate that our proposed method can achieve higher utility than existing multi-dimension LDP aggregation approaches.