Taming false alarms from a domain-unaware c analyzer by a bayesian statistical post analysis

  • Authors:

  • Yungbum Jung;Jaehwang Kim;Jaeho Shin;Kwangkeun Yi

  • Affiliations:

  • Programming Research Laboratory, School of Computer Science and Engineering, Seoul National University;Programming Research Laboratory, School of Computer Science and Engineering, Seoul National University;Programming Research Laboratory, School of Computer Science and Engineering, Seoul National University;Programming Research Laboratory, School of Computer Science and Engineering, Seoul National University

  • Venue:
  • SAS'05 Proceedings of the 12th international conference on Static Analysis
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

We present our experience of combining, in a realistic setting, a static analyzer with a statistical analysis. This combination is in order to reduce the inevitable false alarms from a domain-unaware static analyzer. Our analyzer named Airac(Array Index Range Analyzer for C) collects all the true buffer-overrun points in ANSI C programs. The soundness is maintained, and the analysis' cost-accuracy improvement is achieved by techniques that static analysis community has long accumulated. For still inevitable false alarms (e.g. Airac raised 970 buffer-overrun alarms in commercial C programs of 5.3 million lines and 737 among the 970 alarms were false), which are always apt for particular C programs, we use a statistical post analysis. The statistical analysis, given the analysis results (alarms), sifts out probable false alarms and prioritizes true alarms. It estimates the probability of each alarm being true. The probabilities are used in two ways: 1) only the alarms that have true-alarm probabilities higher than a threshold are reported to the user; 2) the alarms are sorted by the probability before reporting, so that the user can check highly probable errors first. In our experiments with Linux kernel sources, if we set the risk of missing true error is about 3 times greater than false alarming, 74.83% of false alarms could be filtered; only 15.17% of false alarms were mixed up until the user observes 50% of the true alarms.