On-demand time-decaying bloom filters for telemarketer detection

Abstract

Several traffic monitoring applications may benefit from the availability of efficient mechanisms for approximately tracking smoothed time averages rather than raw counts. This paper provides two contributions in this direction. First, our analysis of Time-decaying Bloom filters, formerly proposed data structures devised to perform approximate Exponentially Weighted Moving Averages on streaming data, reveals two major shortcomings: biased estimation when measurements are read in arbitrary time instants, and slow operation resulting from the need to periodically update all the filter's counters at once. We thus propose a new construction, called On-demand Time-decaying Bloom filter, which relies on a continuous-time operation to overcome the accuracy/performance limitations of the original window-based approach. Second, we show how this new technique can be exploited in the design of high performance stream-based monitoring applications, by developing VoIPSTREAM, a proof-of-concept real-time analysis version of a formerly proposed system for telemarketing call detection. Our validation results, carried out over real telephony data, show how VoIPSTREAM closely mimics the feature extraction process and traffic analysis techniques implemented in the offline system, at a significantly higher processing speed, and without requiring any storage of per-user call detail records.

DSPBench: A Suite of Benchmark Applications for Distributed Data Stream Processing Systems

“…The application (called VoipStream or simply VS) detects telemarketing users by analyzing Call Detail Records (CDRs) using a set of filters based on time-decaying bloom filters [82]. A similar application is used in the evaluation of 8 http://www.cse.iitb.ac.in/debs2014/?page_id=42 the BlockMon system [68].…”

Revealing Hidden Hierarchical Heavy Hitters in network traffic

“…As the existence of hidden HHH has been revealed, we need to consider new directions to streaming algorithms which are based on continuous-time operation and can overcome the accuracy limitations of the original disjoint window approaches. More specifically, as a first step towards this evaluation, we consider to implement a Time-decaying Bloom Filter and its extension [2] as a proof of concept. Although we are open to other solutions, we have chosen that particular streaming algorithms for its simplicity.…”

On Moving Averages, Histograms and Time-DependentRates for Online Measurement

“…In [10] we presented a rate measurement method based on exponential smoothing. It was leveraged by the authors in [11] for time-decaying Bloom filters. Furthermore, we introduced the concept of moving histograms in [12] to calculate time-dependent quantiles.…”