|
" An Architecture for Payload Based Network Traffic Classification Via Super Resolution and Transfer Learning "
Muhammad, Waqar
Esposito, Flavio
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1054682
|
| Doc. No
|
:
|
TL53799
|
| Main Entry
|
:
|
Muhammad, Waqar
|
| Title & Author
|
:
|
An Architecture for Payload Based Network Traffic Classification Via Super Resolution and Transfer Learning\ Muhammad, WaqarEsposito, Flavio
|
| College
|
:
|
Saint Louis University
|
| Date
|
:
|
2020
|
| Degree
|
:
|
M.S.
|
| student score
|
:
|
2020
|
| Note
|
:
|
48 p.
|
| Abstract
|
:
|
In modern data centers, a single high-speed computer network generateshundreds of gigabytes of flow statistics per day. These flows are often sampled and analyzed for many network management purposes, such as network debugging, traffic classification, and anomaly detection. Several of these DevOps operations are automated with modern machine learning and deep learning algorithms for faster and more accurate decision making. Regardless of the problem being solved, training any machine learning or deep learning algorithm over a large amount of data may consume a massive amount of resources. In computer network management, it is unfeasible to create these models from scratch for every network topology (the dynamic graph representing the servers and connections) since such topologies are different, and using data or model collected on one network is not always useful for another network. To overcome the problem of re-training, we propose a software architecture to avoid the need for training a machine learning model from scratch for every network topology but instead utilizing the models already created on a massive amount of data and having good performance using transfer learning. In particular, we design architecture to solve the problem of network traffic classification with transfer learning efficiently. Our architecture also includes a super-resolution component to increase the created model's performance by converting the low-resolution input of a small network topology into high-resolution. To test our architecture, we created a traffic matrix from our own generated dataset based on a topology of 50 nodes through the Mininet network emulator. Each router sends a data packet to every other router in the topology. We classified and assigned a label (low-payload) 0 or (high-payload) 1 to each router-to-router communication based on a flow duration at a certain threshold. Our study results indicate that applying transfer learning can increase the accuracy performance of the base model. Still, it can also lead to a severely over-fitted model if we do not have enough training data available.
|
| Descriptor
|
:
|
Computer science
|
|
|
:
|
Educational technology
|
| Added Entry
|
:
|
Esposito, Flavio
|
| Added Entry
|
:
|
Saint Louis University
|
| |