Real Faults in Deep Learning Fault Benchmarks: How Real Are They?

TL;DR

This paper critically examines the realism and reproducibility of faults in DL benchmarks, revealing that only a small fraction are truly representative and reproducible, highlighting challenges in fault evaluation.

Contribution

The study provides a manual analysis of 490 faults, assessing their realism and reproducibility, and exposes limitations in current DL fault benchmarks.

Findings

Only 18.5% of faults meet realism criteria

Faults were reproducible in only 52% of cases

Most faults do not accurately reflect real-world issues

Abstract

As the adoption of Deep Learning (DL) systems continues to rise, an increasing number of approaches are being proposed to test these systems, localise faults within them, and repair those faults. The best attestation of effectiveness for such techniques is an evaluation that showcases their capability to detect, localise and fix real faults. To facilitate these evaluations, the research community has collected multiple benchmarks of real faults in DL systems. In this work, we perform a manual analysis of 490 faults from five different benchmarks and identify that 314 of them are eligible for our study. Our investigation focuses specifically on how well the bugs correspond to the sources they were extracted from, which fault types are represented, and whether the bugs are reproducible. Our findings indicate that only 18.5% of the faults satisfy our realism conditions. Our attempts to reproduce these faults were successful only in 52% of cases.