You are on page 1of 23

Wave Cusum

Marco Antonio Aguilar Cortés


Wave Cusum
Data stream

 Turnstile model
 Let be the input Stream
 Every item in the input stream is a tuple
 Where 𝑖𝑡 is the key (ip address) and 𝑐𝑡 is the weight (number of received bytes)
for such key.
 Turnstile model allows the monitoring of arrivals and departures, it means 𝑐𝑡 can
be either positive or negative.
Hash functions

 Where 𝑎𝑖 is choosen randomly in the set (0,1..,p-1)


 p is an arbitrary prime number (Mersenne number)
 w is set to 512.
Probabilistic data structures
Wavelet Transform

2 Basic
2 Kind of
What does it concepts in Why not
What is? wavelet
works for? wavelet Fourier?
transforms
transform
Wavelet

 Wavelet is a rapidly decaying wave like oscilation that has 0 mean


What does it works for?

 Wavelet transform help us decompose the signal to make them easier to


analyze the changes in it.
2 basic concepts in wavelet transform

Scaling Shifting
Scaling a wavelet

 Scaling is stretching or shrinking the signal in time


 s is a positive scale factor.
Scaling a wavelet
Scaling a wavelet
Shifting

 Shifting means delaying or advancing the wavelet along the signal.


Selected wavelet function

 In this study, reseachers have selected Daubechies-4 wavelet


Discrete wavelet transform
Why not Fourier?

 Fourier
transform works
on the
frequency
domain and
Wavelet
transforms
works on time
domain
Cusum

 What is cusum? Cumulative Sum


 Where 𝜇𝑛 𝑎𝑛𝑑 𝜎𝑛 are the mean value and standard deviation until step n
 𝜇𝑛 𝑎𝑛𝑑 𝜎𝑛 are calculated by using the EWMA (Exponentially Weighted
Moving Average).
 c is a tunable parameter, researchers used 0,5.
 Anomalies are detected when the increment reaches threshold
Test data

 Publicly available Abilene/Internet2 dataset


 2016 files containing each containing a 5 min bin time traffic.
Comparison between signals

Original signal Reconstructed signal


Cusum comparison
Results comparison
Results comparison
Results comparison

You might also like