-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmeeting_notes.txt
56 lines (48 loc) · 3.36 KB
/
meeting_notes.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
6/9/2022 Meeting:
- I need to implement Batch normalisation on my features as the noise affects some features more then others (normalise input features)
- I should implement N fold cross validation, which pretty much makes many training and test sets then takes the average
- I need to create a confusion matrix for each model to view the results
9/8/2022 Meeting:
Questions:
- How deeply should I explore how DP interferes with possible attack vectors?
- Yep this should be done
- H & S assessment for proposal?
- Talk about possible speed bumps in the project
- Uncertainity in plan?
- Try to be realisatic
3/3/2023 Meeting:
- What could be the noviality (What have I done that hasn't been done before)?
- What about the Pipeline?
- Proving how private it is?
- Losses in accuracy when not private (Using epsilon definition to quanify how accuracy changes with respect to privacy)?
- Look for other metrics of privacy?
- Create plan for rest of semester so show to Dan in my next nexting?
- Poster/Demo will be a A3 slide for the poster, which will be about 2 minutes of the demo and the rest of the 30 minutes is for demoing my work
- Dan will send me some previous Thesis some I can get an idea of how to write mine
10/3/2023 Meeting:
- Dan said to go with Extend a DP implementation from literature with some advanced techniques, such as advance composition from RDP, bounding the total privacy cost for epsilon by calculating an epsilon per iteration, etc (Effectively take from DP ML model from a paper and try to improve it)
- Dan said that Thursday morning is a better time for him for this meeting
- Epsilon, Delta Differential Privacy definition is fine for proving my model is private
- Dan seems fine with the idea of 4 weeks for writing the report
16/3/2023 Meeting:
- I told Dan about the LoPub paper I was looking at, and he liked it
- He said to work on providing a summary of the paper and the method, then look at how to improve it
- I've come to the decision that my paper will be focused on Local Differential Privacy with the idea of providing a small improvement on this LoPub method
- I should probably change the name of my paper to mention that LDP aspect
23/3/2023 Meeting:
- I need to create an overall pipe line of the project for showing the dataset, the DP algorithm that adds the noise, the ML algorithm that is trained on the noisy data and the metric by which the privacy is measured.
30/3/2023 Meeting:
- Show the different mechnaisms that can be used (Done)
- Explain how the noise is added (Done)
- Show the relationship of privacy to epsilon and delta
- Line box and whisker plot for accuracy vs privacy (Done)
- Always use 3 dots for ... (Done)
- Send Dan an email asking about access to the cluster for run my experiments (I can manage to get reasonable computation times by using incements of 0.5 for epsilon)
27/4/2023 Meeting:
- Dan said that I was doing the best of most of the thesis students, which helped calm my anxiety
- For the demo, It will be show the poster for a minute or two then go into demoing what I have done
- This could be maybe a high level overview of what the code does since showing the code straight is a little much to take in
- Dan says he can send me a thesis report from a previous student to help with my structure for the report
- I still want to ask him how I should frame my report
4/5/2023 Meeting:
-