Trace Enrichment & Controlled Traffic Generation
Paper (IEEE IPCCC 2025) · Code: not yet public
Traffic classification benchmarks mix many sources of variation at once, so it is impossible to say which factor an augmentation is actually compensating for.
Approach. Build a testbed that isolates one network factor at a time: ten Dockerized services spanning five application classes, from video streaming and conferencing down to email and FTP, run through a controlled workload sweep.
Finding. Augmentation is method-specific. The identical augmentation that improves packet-size-sequence classifiers measurably degrades timestamp-based ones — so the useful question is not whether augmentation helps, but which signal the classifier is reading.
My role: first author — testbed design, augmentation methods, implementation, and evaluation.
Stack: Python, Docker, network traffic capture and analysis.
Links