Internet Traffic Classification Using Bayesian Analysis Techniques Presentation by Umamaheswararao K
Overview Statistical Method Uses Supervised Machine learning Uses only flow records Based on descriminators of the flows - port, inter-packet gap etc… Applies Naïve Bayesian techniques Reasonably high accuracy
Machine Learned Classification Deterministic Approach Assigns data points to one of mutually exclusive classes Probabilistic Approach assigns the flow with probabilties of belonging to certain class - Current technique falls into this category
Probabilistic Approach: Can Identify similar Characteristics of flows after their probabilistic class assignment Robust to measurement error Provides a mechanism for quantifying class assignment probabilities Available in many implementations
Terminology Objects: Entities to be classfied – here traffic-flows which is a tuple of src/dst IP, protocol, src/dst port Discriminators: Characteristics parameterizing the flow behaviour – flow duration, TCP port etc - Here only complete TCP connections are considered
Discriminators/Categories
Analysis Tools Naïve Bayesian Classifier
Bayes Tech: Contd.. Assumptions – Discriminators Independent TCP header length proportional to pak len or vice versa Discriminator distribution is assumed to be normal (Gaussian) - Distribution can be multimodal
Example
Naïve Bayes: Kernel Estimation Descriminator distribution is not Gaussian
Naïve Bayes vs Kernel
Descriminator selection Remove Irrelevant descriminators Cannot differentiate the class Same distribution for all classes Remove Redundant descriminators highly correlated with another discriminator
Descriminator reduction: Filter Uses characteristics of training data to see how relevant the descriminator to the class degree of correlation b/w discriminator & class Wrapper uses results of a classifier to build optimal set
FCBF Fast-correlation based filter for discriminator filtering Two stage process Identifying the relevance of a discriminator Identifying the redundancy of a feature with respect to discriminators
Results
Results: contd.. Accuracy: Correctly classified flows/Total number of flows Trust: Probability that a flow that has been classified into some class in fact from this class
Naïve Bayes- Trust
Trust: Kernel est.
Results for new data set
Identification of discriminators
Strengths Payload access not needed High accuracy and Trust with FCBF Easily implementable Single flow based (a strength and a weakness) Allows any categorization
Weaknesses Bunch of them but then …? Accuracy/Trust depends mainly on how good the training set is Trust of some classes is really poor works on flow based, characterization some flows require to see many flows (eg. Attacks) Temporal stability is not really good Discriminators are dependent on network dynamics
Weaknesses: Contd… Training is not automatic Assumes discriminator independence Gaussian distribution assumption inaccurate
Future Work A significantly new approach hence can lead to many ideas Spatial independence of traffic classification Check from weaknesses section