Android Memory Leak Benchmark
Large-scale, manually validated benchmark of 650+ Android projects with confirmed memory leak patterns, built using Soot and FlowDroid.
A core component of my dissertation at the Software Analysis and Intelligence Lab, Wichita State University. Advised by Dr. Zhiyong Shan.
Problem. Existing Android static analysis benchmarks are small, synthetically generated, or lack ground-truth validation — making it difficult to evaluate whether new tools represent genuine improvements.
Pipeline Overview
Stage 1 — Candidate Collection. Mined 650+ open-source Android repositories from GitHub, filtering by activity lifecycle complexity and minimum build health criteria.
Stage 2 — Static Analysis Pass. Ran FlowDroid (built on Soot) with a custom source/sink configuration targeting Context and View objects held past their expected lifecycle boundaries.
Stage 3 — Manual Triage. Reviewed each flagged case against the source code and cross-referenced with issue trackers to confirm ground truth.
Benchmark Statistics
| Metric | Value |
|---|---|
| Android projects analyzed | 650+ |
| Confirmed leak patterns | In progress |
| Analysis framework | Soot + FlowDroid |
| Ground-truth method | Manual triage + issue tracker |
| HPC cluster | NSF Expanse (Slurm) |
| Programming language | Java (Android), Python (tooling) |
Why Manual Triage Matters
Without manual validation, a benchmark reflects the tool’s false positive rate rather than the true distribution of Android memory bugs. Our triage step ensures every entry in the dataset has a confirmed ground-truth label, making it suitable for training and evaluating AI-based bug detection models.
Stack. Java · Soot · FlowDroid · Android SDK · Python · Slurm (HPC)
Status. Active — benchmark and tooling will be released publicly alongside the dissertation paper.