JITI
Adaptive ensemble classification for network traffic identification

Overview
JITI (also known as AC-DC: Adaptive Constraint-Driven Classification) is a framework for adaptive ensemble classification that intelligently balances accuracy and efficiency in network traffic identification by dynamically selecting between multiple classifiers based on system conditions.
The Problem
Network traffic classifiers face a fundamental trade-off:
- Flow-based classifiers using pre-computed statistics are fast but less accurate
- Packet-capture classifiers are more accurate but slower and memory-intensive
Existing approaches commit to a single classification method, forcing operators to choose between accuracy and efficiency.
What JITI Does
- Adaptive Scheduling: Dynamically tracks system memory availability and incoming traffic rates
- Intelligent Classifier Selection: Determines the optimal classifier and batch size based on current constraints
- Ensemble Approach: Maintains a pool of classifiers with different performance characteristics
- Real-time Adaptation: Switches between classifiers as system conditions change
Key Features
- Dynamic Resource Management: Monitors memory and CPU availability in real-time
- Performance Optimization: Selects the best classifier for current system state
- Traffic-Aware Processing: Adapts to varying network traffic loads
- Batch Size Optimization: Adjusts processing batch sizes to maximize throughput
Use Cases
- Network traffic prioritization
- Anomaly detection
- Service identification
- Real-time traffic classification under resource constraints
Results
- 100%+ improvement in classification performance vs. flow-statistics-only approaches
- Comparable accuracy to state-of-the-art packet-capture classifiers (less than 12.3% lower F1-Score)
- 150x faster processing than complex packet-capture methods
Resources
Citation
@article{jiang2023acdc,
title={AC-DC: Adaptive Ensemble Classification for Network Traffic Identification},
author={Jiang, Xi and Liu, Shinan and Naama, Saloua and Bronzino, Francesco and Schmitt, Paul and Feamster, Nick},
journal={arXiv preprint arXiv:2302.11718},
year={2023}
}