Download Poster_ppt_version - Simon Fraser University

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Recursive InterNetwork Architecture (RINA) wikipedia , lookup

Airborne Networking wikipedia , lookup

IEEE 1355 wikipedia , lookup

Peering wikipedia , lookup

Routing in delay-tolerant networking wikipedia , lookup

Transcript
Detecting BGP Anomalies
Using Machine Learning Techniques
Student: Zhida Li. Instructor: Steve Whitmore.
Simon Fraser University, Vancouver, British Columbia, Canada
BORDER GATEWAY PROTOCOL
• Border Gateway Protocol (BGP) is an interdomain routing protocol used in
networks consisting of a large number of Autonomous Systems (ASs).
• Propagation of the BGP routing information is susceptible to misconfigurations,
power outages, malicious attacks, and worms.
• The main function of BGP is to select the best routes between ASes based on
routing algorithms and network policies.
• Determining the anomalies and their causes is useful for assessing loss of data
and connectivity.
LSTM
A = LSTM cell
PERFORMANCE EVALUATION
• SVM performance:
ACCURACY AND F-SCORE USING THE SVM 2 MODELS FOR UNBALANCED AND BALANCED DATASETS
Xt = (fearture1, …, feature n)
ht = binary label for Xt
ht=1 if anomaly
ht= 1 otherwise.
• BGP anomaly detection system design relies on machine learning techniques.
• LSTM performance:
• We use well-known classifiers and exploit their ability to reliably detect
network anomalies in datasets of known BGP network anomalies.
ACCURACY AND F-SCORE USING THE LSTM MODELS FOR UNBALANCED AND BALANCED DATASETS
BGP DATASETS
• Analyzed Internet routing data are acquired from two projects that provide
valuable information to networking research:
• Routing Information Service (RIS) project initiated in 2001 by the Réseaux
IP Européens (RIPE) Network Coordination Centre (NCC)
• These projects collect and store routing data that provide a unique view of
the Internet topology.
• Anomalous events considered in this project:
EXPERIMENTAL PROCEDURE
• Step1:
Train and test the three SVM and LSTM models using 37 features.
• Step2:
Select the 10 most relevant features using the three feature selection
algorithms: MID, MIQ, and MIBASE. Train and test the three SVM models
using datasets with and without anomalies. Skip this Step for generating
the LSTM models.
• Step3:
Evaluate the SVM and LSTM models using the accuracy and F-score
measures.
APPROACHES
In this project we attempt to detect various BGP anomalies by applying:
• Support Vector Machine (SVM) models, and
• Step4:
Tune the SVM and LSTM model parameters to achieve the best
performance.
PERFORMANCE MEASURES
• minimum Redundancy Maximum Relevance (mRMR)
Mutual Information Difference (MID)
Mutual Information Quotient (MIQ)
Mutual Information Base (MIBASE).
THE BEST ACCURACY AND F-SCORE OF SVM AND
LSTM MODELS
CONCLUSION
• Feature selection and classification algorithms were used to detect BGP
anomalies.
• The SVM 2 models based on the combination of the Slammer and Code Red I
training datasets achieve better accuracy and F-score than results reported.
• The SVM classifier achieved the highest F-score using balanced datasets. In case
of the unbalanced datasets, the accuracy is higher due to the large number of
the regular testing data.
• Using the SVM classifier may be a feasible approach for detecting BGP
anomalies in communication networks.
REFERENCES
• A Long Short-Term Memory (LSTM) recurrent neural networks.
We use feature scoring algorithms for SVM to select the most relevant features:
• Performance comparison:
• Sensitivity: ratio of identified anomalies (TP) and all labeled anomalies
(true).
• Precision: ratio of identified anomalies (TP) and all data points identified as
anomalous.
• TP: number of anomalous training data points classified as anomaly
• FP: number of regular training data points classified as anomaly
• FN: number of anomalous training data points classified as regular
• TN: number of regular training data points classified as regular.
• N. Al-Rousan and Lj. Trajković, “Machine learning models for classification of BGP anomalies,” in
Proc. 13th IEEE Int. Conf. High Performance Switching and Routing, Belgrade, Serbia, June
2012, pp. 103–108.
• N. Al-Rousan, S. Haeri, and Lj. Trajković, “Feature selection for classification of BGP anomalies
using Bayes models,” in Proc. Int. Conf. Mach. Learning Cybern., Xi'an, China, July 2012, pp.
140–147.
• M. Ćosović, S. Obradović, and Lj. Trajković, “Performance evaluation of BGP anomaly classifiers,”
in Proc. Int. Conf. on Digital Inform., Networking and Wireless Commun., Moscow, Russia, Feb.
2015, pp. 115–120.
• Y. Li, H. J. Xing, Q. Hua, X.-Z. Wang, P. Batta, S. Haeri, and Lj. Trajković, “Classification of BGP
anomalies using decision trees and fuzzy rough sets,” in Proc. IEEE Int. Conf. Syst., Man,
Cybern., San Diego, CA, USA, Oct. 2014, pp. 1331–1336.
ENSC 803
Writing for Publication
August 2016, Burnaby, Canada