This work was undertaken as part of my master thesis at ETH Zurich, from Feb 2022 to Aug 2022, titled Advancing packet-level traffic predictions with Transformers
. We present a new transformer-based architecture, to learn network dynamics from packet traces.
We design a pre-training
phase, where we learn fundamental network dynamics. Following this, we have a fine-tuning
phase, on different network tasks, and demonstrate that pre-training well leads to generalization to multiple fine-tuning tasks.
The experiments conducted in this project are very involved. Understanding and reproducing them from just the code and comments alone will be quite hard, inspite of the instructions mentioned in the given README
. For more detailed understanding, we invite you to read the thesis (direct link
). You can also check out an overview on the presentation slides (direct link
)
For any further questions or to discuss related research ideas, please feel free to contact me by email.
Some results from the thesis have been written as a paper titled A new hope for network model generalization
and the same has been accepted for presentation at ACM HotNets 2022. The paper is now online and open-access, it can be accessed via the ACM
Digital Library via this link, DOI is: 10.1145/3563766.3564104.
The thesis has now been published under the ETH Research collection, which is open access. It can be accessed from here (direct link
)