Transcript A Signal Analysis of Network Traffic Anomalies Paul Barford

A Signal Analysis of Network
Traffic Anomalies
Paul Barford
with Jeffery Kline, David Plonka, Amos Ron
University of Wisconsin – Madison
Summer, 2002
Motivation
• Traffic anomalies are a fact of life in computer networks
– Outages, attacks, etc…
• Anomaly detection and identification is challenging
– Operators typically monitor by eye using SNMP or IP flows
• Obviously, this does not scale!
– Simple thresholding is ineffective
– Some anomalies are obvious, other are not
• Characteristics of anomalous behavior in IP traffic are
not well understood
– Do same types of anomalies have same characteristics?
– Can characteristics be effectively used in detection systems?
2
Introduction
• Objective: Improve our understanding network traffic
anomalies
• Approach: Wavelet analysis of data set that includes IP
flow data, SNMP data and a catalog of observed
anomalies
• Method: Integrated Measurement Analysis Platform for
Internet Traffic (IMAPIT)
• Results: We demonstrate how anomalies can be
exposed using wavelets and develop new method for
exposing short-lived events
3
Related Work
• Network traffic characterization
– Eg. Caceres89, Leland93, Paxson97, Zhang01
• Focus on typical behavior
– Abry98 use wavelets to analyze LRD traffic
• Fault and anomaly detection techniques
– Eg. Feather93, Brutlag00
• Focus on thresholds and time series models
– Eg. Paxson99
• Rule based tool for intrusion detection
– Eg. Moore01
• Backscatter technique can be used to identify DoS attacks
– Eg. Huang01
• Wavelet-based approach to detecting network performance problems
4
Simple Network Management Protocol
• SNMP is the standard protocol for monitoring/managing
networked systems
• SNMP defines a set of MIB (management information
base) data exported from routers
– RFC2863
• We sample High Capacity Interface using MRTG (MultiRouter Traffic Grapher) at 5 minute intervals
– Archive byte and packet traffic in each direction
– 64-bit counters on each of 15 WAN links
• SNMP count precision is yet to be determined…
5
IP Flows
• An IP Flow is defined as a unidirectional series of
packets between source/dest IP/port pair over a period of
time
{SRC_IP/Port,DST_IP/Port,Pkts,Bytes,Start/End Time,TCP Flags,IP Prot …}
– Exported by Lightweight Flow Accounting Protocol (LFAP)
enabled routers (Cisco’s NetFlow, Juniper cflowd flow export)
• We use FlowScan [Plonka00] to collect and post-process
IP flow data collected at 5 minute intervals
– Combines flow collection engine, database, visulaization tool
– Provides a near real-time visualization of network traffic
– Breaks down traffic into well known service or application
6
7
Our Approach to Data Gathering
• Consider anomalies in IP flow and SNMP data
– Collected at UW border router (Juniper M10)
– Archive of ~6 months worth of data (packets, bytes, flows)
– Includes catalog of anomalies (after-the-fact analysis)
• Group observed anomalies into four categories
– Network anomalies (41)
• Steep drop offs in service followed by quick return to normal behavior
– Flash crowd anomalies (4)
• Steep increase in service followed by slow return to normal behavior
– Attack anomalies (46)
• Steep increase in flows in one direction followed by quick return to normal
behavior
– Measurement anomalies (18)
• Short-lived anomalies which are not network anomalies or attacks
8
Our Approach to Analysis
• Wavelets provide a means for describing time series
data that considers both frequency and time
– Particularly useful for characterizing data with sharp
spikes and discontinuities
• More robust than Fourier analysis which only shows what
frequencies exist in a signal
– Tricky to determine which wavelets provide best
resolution of signals in data
• We use tools developed at UW which together make
up IMAPIT
– FlowScan software
– The IDR Framenet software
9
Our Wavelet System
• After evaluating different candidates we selected a
wavelet system called Pseudo Splines(4,1) Type 2.
– A framelet system developed by Daubechies et al. ‘00
– Very good frequency localization properties
• Three output signals are extracted from input
– Low Frequency (L): synthesis of all wavelet coefficients
from level 9 and up
– Mid Frequency (M): synthesis of wavelet coefficients 6, 7, 8
– High Frequency (H): synthesis of wavelet coefficients 1 to 5
• Thresholding (set to zero all coefficients whose absolute value is below
a threshold) is used on these coefficients
10
Ambient IP Flow Traffic
11
Ambient SNMP Traffic
12
Byte Traffic for Flash Crowd
13
Average Packet Size for Flash Crowd
14
Flow Traffic During DoS Attacks
15
Byte Traffic During Measurement Anomalies
16
Anomaly Detection via Deviation Score
•
We develop an automated means for identifying shortlived anomalies based on variability in H and M signals
1. Compute local variability (using specified window) of H and
M parts of signal
2. Combine local variability of H and M signals (using a
weighted sum) and normalize by total variability to get
deviation score V
3. Apply threshold to V then measure peaks
•
Our analysis shows that V peaks over 2.0 indicate
short-lived anomalies with high confidence
–
We threshold at V = 1.25 and set window size to ~3 hours
17
Deviation Score for Three Anomalies
18
Deviation Score for Network Outage
19
Anomalies in Aggregate Signals
20
Hidden Anomalies in Low Frequency
21
Deviation Score Evaluation
• How effective is deviation score at detecting anomalies?
– Compare versus set of 39 anomalies
• Set is unlikely to be complete so we don’t treat false-positives
– Compare versus Holt-Winters Forecasting
• Sophisticated time series technique
• Requires some configuration
• Holt-Winters reported many more positives and sometimes
oscillated between values
Total
Candidate
Anomalies
Candidates
detected by
Deviation
Score
Candidates
detected by
Holt-Winters
39
38
37
22
Conclusion and Next Steps
• We present an evaluation of signal characteristics of network
traffic anomalies
– Using IP flow and SNMP data collected at UW border router
• 106 anomalies have been grouped into four categories
– IMAPIT developed to apply wavelet analysis to data
– Deviation score developed to automate anomaly detection
• Results
– Characteristics of anomalies exposed using different filters
and data
– Deviation score is effective detection method
• Future
– Development of anomaly classification methods
– Application of results in (distributed) detection systems
23