基于二值化网络的学习型布隆过滤器优化研究

Research on optimization method of learned Bloom Filter based on binarized neural network

学习型布隆过滤器的查询速度比布隆过滤器慢一个数量级,这阻碍了学习型布隆过滤器在对实时性要求较高场合中的应用。文中提出了一种基于二值化网络的学习型布隆过滤器优化方法。将学习型布隆过滤器中预过滤器的权重和激活进行二值化,可以加快学习型布隆过滤器的查询速度。使用恶意和良性网址数据集测试了不同假正例率下二值化学习型布隆过滤器和学习型布隆过滤器的总体空间占用和查询时间。为了对查询速度效果改善有直观的认识,在不同平台上分别进行了实验。结果显示,二值化学习型布隆过滤器的查询速度是学习型布隆过滤器的1.5~2倍。

The query speed of the learned Bloom filter is an order of magnitude slower than that of the Bloom Filter,which hinders the application of the learned Bloom Filter in occasions with high requirements for real⁃time performance.In this paper,a learned Bloom Filter optimization method based on binarized networks was proposed.The query speed of the learned Bloom Filter can be accelerated by binarizing the weights and activations of the pre⁃filter in the learned Bloom Filter.The overall space occupation and query time of the binarized learned Bloom Filter and the learned Bloom Filter with different false positive rates are tested using malicious and benign URL dataset.In order to have an intuitive understanding of the query speed effect improvement,separate experiments were conducted on different platforms.The results show that the query speed of the binarized learned Bloom Filter is 1.5 to 2 times faster than that of the learned Bloom Filter.