uvm-plaid / programming-dp Goto Github PK

There is a display of three records just prior to the text:

This time, there are three rows returned - and we don’t know which one is the real Karrie.

The three records have the same date of birth ("9/7/1967") but different ages. This is impossible.

At the bottom of the page in the code snippets that call sparse() and range_query_svt(), it shows it as throwing an error.

When you define isKAnonymized in chapter 3, the query assumes the numbers to be strings, not ints. This causes it to fail to detect anything in your examples.

The formula $\sigma^2 = 2 s^2 ln(1.25/\delta) / \epsilon^2$ is not valid for every epsilon / delta. Theorem 3.22 in Dwork, Roth is only for $\epsilon \in (0, 1)$. See theorem 4 in Balle et al (2018) showing this bound does not hold for $\epsilon > 1$.

It would probably be good to add a caveat that the bounds in the book are not valid for every $\epsilon$, in a footnote or something.

I could not locate any documentation for the data. Is there any?

Hey, thanks for this great page. I noticed that you host the static page on GitHub Pages, but don't have the required AAAA DNS records to make it accessible over IPv6. The good news is – You just have to add those to the domain and you have full IPv6 support.
Just follow the official guide from GitHub: https://docs.github.com/en/pages/configuring-a-custom-domain-for-your-github-pages-site/managing-a-custom-domain-for-your-github-pages-site#configuring-an-apex-domain

In Chapter 4, in the Sequential Composition section between In[4] and In[5], 'likelihood' is misspelled 'liklihood'.

I noticed a typo in the PDF downloaded from programming-dp.com/book.pdf that seems to be fixed already in the repo. Cell 75 is ch2.ipynb outputs the result False, but the pdf says True on page 18.

I appreciate the book so I wanted to point out this minor discrepancy.

Fix the SAA graphs: they need legends and color disambiguation

This distinction is important and the lack of clear discussion creates lots of confusion later on.

Hi. In equation 36 of the Machine Learning Section of the book and the subsequent paragraph, the clipped gradient is written as

when it really should be written as

The second term is missing a zero.

(I wrote this latex in this github issue using this trick here.)

Add a checklist for privacy analysis

I have found a typo in the De-identification chapter. In the section describing re-identifying Karrie Trusslove's data, the word "information" is spelled "informatino".

Full line is:

• We can look at the differences between the rows to determine what additional auxiliary informatino would help us to distinguish them (e.g. sex, occupation, marital status)

In the Machine Learning chapter, there are some typos in the gradient clipping section, and the explanation of sensitivity is insufficient.

In "Variants of Differential Privacy" section, when calculating noises_seq for advanced composition why
noises_seq = [16knp.log(1.25/delta)np.log(1/delta)/(epsilon**2) for k in ks]
is used instead of
noises_seq = [16knp.log(2.5k/delta)*np.log(2/delta)/(epsilon**2) for k in ks] as derived earlier?

Equation (2) in chapter 3 - Laplace Mechanism should use \left and \right for braces on the right side:

\begin{equation}
F(x) = f(x) + \textsf{Lap}\left(\frac{s}{\epsilon}\right)
\end{equation}

Before:

After:

There are also other equations which would look a lot better if the \left and \right were used for braces:

• Chapter 5 - equation 3
• Chapter 7 - equation 18
• Chapter 8 - eq 22, 23, 25
• Chapter 9 - eq 30, 31, 32
• Chapter 13 - eq 38

Whether $\delta \ge 0$ in the previous line of formula (12) should be $\delta' \ge 0$ ?

The section on advanced composition has a line that says:

" for \epsilon', \delta \ge 0, the total privacy cost of the entire k-fold adaptive composition is equal to \epsilon', \delta, where <formula for \epsilon'>. "

The \epsilon' should be outside the \forall quantifier right? My understanding was that the order of quantifier goes "\forall \epsilon, \delta, \exists \epsilon' given by formula", but I might be misunderstanding something. Specifically, I think the dash should be removed in the first \epsilon'.

Not as important, but I think there should be brackets around the \epsilon', \delta at the end of the sentence?

...adult_with_pii.csv dataset.

This was seen and tested in a lecture exercise. You receive a KeyError on the first Education key found.

Thank you to Mohammad Moghimi for this suggestion.

Chapter 2 --
Informally, we say that a dataset is “k-Anonyized” -> Informally, we say that a dataset is “k-Anonymized”

Chapter 6 --
this is the “Euclidian distance,” -> this is the “Euclidean distance”,

Chapter 6 --
which we will call the catastrophe mechansim -> which we will call the catastrophe mechanism
fails gracefully, rather than catistrophically -> fails gracefully, rather than catastrophically

There is a duplicated word on the line:

Informally, we say that that a dataset is “k-Anonyized” for a particular k if each...

In subsection "Is Karrie Special?", the "A good way to guage the effectiveness of this type of attack is to look at how “selective” certain pieces of data are.", the "guage" should be "gauge".

Add a paragraph on the homogeneity attack in the k-anonymity chapter.

https://uvm-plaid.github.io/programming-dp/notebooks/ch6.html#advanced-composition –
the total privacy cost under both sequential compsosition -> the total privacy cost under both sequential composition

https://uvm-plaid.github.io/programming-dp/notebooks/ch7.html#sample-and-aggregate –
In this simple instantation -> In this simple instantiation

https://uvm-plaid.github.io/programming-dp/notebooks/ch8.html#variants-of-differential-privacy –
eliminating the catastropic failure mode -> eliminating the catastrophic failure mode