Bidirectional Anonymous Private Set Intersection Based on Obfuscation

As a key pre-step of vertical federated learning, private set intersection (PSI) technology finds the intersection of data from each participant without revealing their unique samples. However, the current PSI scheme incurs significant transmission costs, and participants can infer what data the other party possesses through the intersection set, violating anonymization requirements set by laws and regulations. To fill these gaps, we analyze the advantages and disadvantages of the existing PSI schemes based on the basic principles of different fundamental disciplines in the field of PSI research, and opt to construct a PSI scheme based on the oblivious transfer protocol and its extension, the oblivious pseudo random function. Subsequently, we improve the constructed scheme with cuckoo hash to address the transmission cost problem, reducing the complexity of communication transmission overhead from $O\left(n^{2}\right)$ to $O(n)$. Finally, in view of the lack of meeting the anonymization requirement in the field of PSI, we construct a bidirectional anonymization scheme by means of obfuscation expansion, and use the method of controlling variables to derive the effect of different variable variations on the final modeling accuracy through a large number of randomization experiments. Experimental results show that the proposed scheme can achieve bidirectional anonymous PSI while maintaining the performance of the federated model.