Low-Complexity and Secure Clustering-Based Similarity Detection for Private Files

Detection of the similarity between files is a requirement for many practical applications, such as copyright protection, file management, plagiarism detection, and detecting duplicate submissions of scientific articles to multiple journals or conferences. Existing methods have not taken into consideration file privacy, which prevents their use in many delicate situations, for example when comparing two intellectual agencies' files where files are meant to be secured, to find file similarities. Over the last few years, encryption protocols have been developed with the aim of detecting similar files without compromising privacy. However, existing protocols tend to leak important data, and do not have low complexity costs. This paper addresses the issue of computing the similarity between two file collections belonging to two entities who desire to keep their contents private. We propose a clustering-based approach that achieves 90% accuracy while significantly reducing the execution time. The protocols presented in this study are much more efficient than other secure protocols, and the alternatives are slower in terms of similarity detection for large file sets. Our system achieves a high level of security by using a vector space model to convert the files into vectors and by applying Paillier encryption to encrypt the elements of the vector separately, to protect privacy. The study uses the application of the Porter algorithm to the vocabulary set. Using a secure cosine similarity approach, a score for similar files was identified and the index of the similarity scores is returned to the other party, rather than the similar files themselves. The system is strengthened by using clustering for files, based on the k-means clustering technique, which makes it more efficient for large file sets.