[Overview] Chapter 1

[bartvanb]

Desired solution

TODO:

• [x] #89
• [x] #60
• [x] Key points need to be added #58
• [x] Questions still need to be done #58
• [x] "Only when all parties work together, the original value can be retrieved." this is not true. This is n-1 parties work together. - #63
• [x] Differential Privacy: Now the description is a bit vague. I would mention that in practice this means adding noise to the data whilst maintaining statistical properties when the data set used is large enough

  • 63

• [x] Blocks upon blocks: I like this section. Please also add that differential privacy can be applied to the datasets before they are used by FL, MPC or HE.

  • 63

• [x] Data partitioning: A picture would be really helpful here. #49
• [x] Data partitioning: "Data can even be horizontally and vertically partitioned at the same time. In these cases, it may be necessary to combine multiple techniques." I've never seen this? Could you give an example? #63
• [x] Include #49

Discuss:

• [x] "There are different ways in which privacy risks can be mitigated. We will focus on the idea of federated analysis." -> maybe very briefly mention other ways with a link for people that are interested?

  • 63

• [x] #94
• [x] Can we make a simple example on how differencing works in the Federated Learning section? For example lets say we have Alice Bob, Clair and Dave in our dataset.
  • We have a column age.
  • We have an average algorithm
    • The algorithm can make sub selection of the data based on age. For example age > 30.
  • Create a protocol to obtain the age of each individual in the dataset

• 63

• [ ] #139
  • Not necessarily high priority for me
• [x] "This is why there are privacy regulations in place like GDPR to prohibit easy access to this wealth of data." -> Maybe also mention the [EHDS](Europese ruimte voor gezondheidsgegevens - Europese Commissie (europa.eu))

  • 63

• [ ] #62 : Not gonna do this. Too much theory
  • Reconstruction
  • Differencing
  • Deep Leakage from Gradients (DLG)
  • Generative Adversarial Networks (GAN)
  • Model Inversion
  • Watermark Attacks
• [x] Maybe we should also mention that we can use FL for classical statistical models.
  • See my reply
• [x] MPC: It only works when parties > 2. So maybe we could use that in an example? #93
• [x] Technology doesn’t solve everything: Maybe add a link to our DPIA?: Data Protection Impact Assessment (DPIA) (vantage6.ai)

  • 63

• [x] Technology doesn’t solve everything: I am missing the data harmonization part in this chapter. This has typically be done before we start any analysis. Techniques like OMOP CDM, FHIR using SNOWMED CT or any common data model can be helpful for this purpose.

  • 63

• [x] Add future changes to chapters: don't think this is relevant

[dsmits]

@bartvanb what do you mean by saying you can use FL for classicic statistical models? I'm currently making the distinction between federated learning and federated analysis. Isn't that enough?

[dsmits]

@bartvanb I made a separate issue about the #62 but I'm kind of thinking not to do it. This chapter is by no means meant to be comprehensive. I did add a differencing attack in #63 (kind of).

I could add some links so people can read up on other attacks though

[bartvanb]

@dsmits I think these were comments from @frankcorneliusmartin (see https://github.com/vantage6/vantage6-workshop/pull/37 for details) - maybe you can discuss what is meant?

(FYI: the list of items above was a concatenation of all findings of Hector, Frank and myself during the full review, combined with everything that was already open issue for this chapter)

[dsmits]

Ok I think this is done now.