SC²S Colloquium - January 19, 2017
|Date:||January 19, 2018|
|Time:||3:00 pm, s.t.|
James W. Browne: Variational Inference for Unsupervised Reverse Engineering of Protocols from Bitstreams
Variational Bayesian methods have recently emerged as leading unsupervised learning methods. With regard to time series, the Deep Variational Bayes Filter (DVBF) has de- livered state of the art performance in extracting system dynamics in an unsupervised fashion. This work attempts to apply the DVBF to bit level signals to extract the un- derlying protocol or at least transmitted information. Key contributions are a modified version with a Long Short Term Memory (LSTM) based transition function to deal with longer signals; as well as an approximately discrete DVBF inspired by the Finite State Transducer (FST) using the Concrete or Gumbel-Softmax distribution for hidden vari- ables. Although the modifications are able to learn the protocol to some extent, neither shows a clear advantage over the original DVBF architecture. The multiple levels of ab- straction that separate bitstreams from the underlying information make it difficult for an unsupervised learning algorithm to extract the true system dynamics.