grizzly/deneir
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

add some additions ideas

Browse source
This commit is contained in:
Arthur Grisel-Davy 2024-01-11 14:58:04 -05:00
parent e30a119529
commit 8471197d4d
1 changed files with 31 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

36
DSD/journal/additions.md
View file

@ -1,11 +1,37 @@
vim:ft=markdown
# Miscelaneous
= Changes
For ACM, the title should not be the same as for the conference paper.
== Multi-shot Classification
# Explanation Additions
These are futher explanations of concepts that were only superficially presentedin the original paper.
## Multi-shot Classification
The system is designed to get the most out of one-shot training data. We explain in the discussion that there is a natural extension to multi-shot training data. However we don't provide much more explanation nor an evaluation of the performances.
-> Explain in details how the algorithm work with multiple training samples per class. Explain that this is a competitive system where samples of the same class compete against each other but play for the same team. There may be a change to make on the computation of the threshold that should only consider distances between different classes (min of distance between two samples of different classes, this has a specific name in hierarchical clustering design).
Provide an evaluation in the case where a pattern can have multiple modes.
* Explain in details how the algorithm work with multiple training samples per class. Explain that this is a competitive system where samples of the same class compete against each other but play for the same team. There may be a change to make on the computation of the threshold that should only consider distances between different classes (min of distance between two samples of different classes, this has a specific name in hierarchical clustering design).
* Provide an evaluation in the case where a pattern can have multiple modes.
## Time Efficiency & Parallelisation
In the discussion we say that the time efficiency of MAD, while not being properly evaluated, is good enougth to not be prohibitive. However this is a very hand-waving argument and a proper evaluation would be more convincing. We also say that the algorithm can be naturally paralelized. This is true but we do not provide more explanation.
* Do a proper time efficiency evaluation of the algorithm. This might be tricky but you took courses about that so figure it out.
* Provide a better explanation of the parallelisation of the algorithm. Provide a pseudo code on the same format as the main pseudo code in the paper, where the // sections are highlighted and explain. If // changes the time efficiency, provide a new evaluation for the // version.
# Novel Additions
These are additions that introduces completely new information that was no in the original paper.
## New Experiment and AC current
The main experiment of the QRS paper (Case Study 2) presented the detection of states for enforcing security rules.
This was great but limited to only one machine with one set of states/rules.
It would be interesting to add a new case study that illustrate the potential with other machines/scenarios.
* Perform an experiment similar to Case Study 2 but focusing on AC current and non-intrusive measurement of power.
* Equip a tower PC with a power clamp on the AC line and perform the same kind of security policy enforcement.
* This might be tricky because all 8 clamps are in use for the LV experiment. Figure it out...
* Perform a case study on a different device than a computer. It needs to be a machine with possible attacks and enougth different states to be interesting. Possible machines could be:
* WAP if there are more states than ON and OFF.
* Smart TV. Could identify getting into menues, playing of specific ports/peripherals, connection status etc. The policies would be related to usage state and not to attacks.