Triangle counting is essential for analyzing network structures and optimizing recommendation systems, yet it can lead to privacy breaches if individual data is not adequately protected during the analysis. Differential privacy has become a widely adopted standard to safeguard personal privacy. Current research mainly focuses on central and local models, which differ in applicability and performance. The central model cannot be applied when graph data is distributed among multiple parties without a trusted central server, while the local model provides unsatisfactory performance. In this paper, we explore a two-party scenario where each party holds private information about a group of users and is not allowed to disclose this information to the other party. We have proposed a scheme called HTTC-DPk, which ensures both differential privacy and k-anonymity, and is better suited to the constraints of the two-party setting compared to both central and local models. Our method integrates the noisy maximum degree computation for both intra-party and inter-party edges, along with the differentially private inter-party triangle counting based on two-party interactions. Additionally, we introduce an enhanced scheme HTTC-DPk* that strikes a balance between accuracy and communication costs, particularly suitable for large-scale graphs. We have provided comprehensive theoretical proof of our scheme's privacy and demonstrated through extensive experiments that our approach performs well under various cases.

Frequent itemset mining is an essential task in data analysis. Therefore, it is crucial to design practical methods for privacy-preserving frequent itemset mining, enabling private data analysis. For two-party data analysis tasks, each party possesses its portion of the data and is reluctant to share the data with the other. While secure computation can enable two-party frequent itemset mining, the output of exact top-$k$itemsets may still leave the adversary a chance to infer the sensitive information. Differential privacy has been utilized in various data analysis tasks to safeguard participating individuals. However, addressing how to ensure differential privacy for two-party frequent itemset mining remains unexplored. To prevent each party’s data from being leaked to the other while achieving differential privacy for releasing the output, this paper investigates the problem of differentially private two-party frequent itemset mining. We first propose a practical method that can efficiently select the frequent items of the union of two confidential databases in a differentially private way without the need to combine all elements. Then we extend this technique for general frequent itemset mining. Extensive experiments were conducted on real-world datasets, and the results show that the proposed method can achieve satisfactory utility with affordable overheads.

Various data analysis tasks take frequent item mining as an essential part of them. Thus, it is crucial to design practical privacy-preserving frequent item mining methods such that private data analysis can be enabled. For two-party data analysis tasks, each party possesses its portion of the data and is reluctant to share the data with the other. Although secure computation can enable two-party frequent item mining, the output of exact top-$k$items may still leave the adversary a chance to infer the sensitive information. Differential privacy has been utilized in various data analysis tasks to protect participating individuals, but how to provide differential privacy for two-party frequent item mining has not been addressed. To prevent each party's data from being leaked to the other while achieving differential privacy for releasing the output, we study the problem of differentially private two-party frequent item mining in this paper. We have proposed a practical protocol that can efficiently select the frequent items of the union of two confidential databases in a differentially private way but does not need to combine all the elements. We have conducted extensive experiments to evaluate the proposed method on real-world datasets, and the results show that it can achieve satisfactory utility with affordable overheads.