BARS: Local Robustness Certification for Traffic Analysis
BARS: Local Robustness Certification for Deep Learning based Traffic Analysis Systems (NDSS'23)
Introduction
BARS is a general local robustness certification framework for Deep Learning (DL) based traffic analysis systems based on Boundary Adaptive Randomized Smoothing. Against adversarial perturbations, local robustness certification is used to certify whether a DL-based model is robust in the neighborhood of a sample.
BARS optimizes the smoothing noise of randomized smoothing and provide tighter robustness guarantee for the traffic analysis domain. The pipeline includes four steps:
- Build Distribution Transformer.
- Optimize the noise shape.
- Optimize the noise scale.
- Certify robustness.
BARS supports three traffic analysis systems:
- Zero-Positive Network Intrusion Detection System, Kitsune (NDSS'18)
- Concept Drift Detection System, CADE (USENIX Security'21)
- Supervised Multi-Classification System, ACID (INFOCOM'21)
Quick Start
1. Environmental Setup
- Basic BARS:
pip install -r requirement_bars.txt - Smoothed Kitsune:
pip install -r requirement_kitsune.txt - Smoothed CADE:
pip install -r requirement_cade.txt - Smoothed ACID:
pip install -r requirement_acid.txt
2. Running Program
-
Please run
python main.py. -
Program arguments are set at the beginning of
main.py.
Citation
TBA.