Questions tagged [differential-privacy]
The differential-privacy tag has no usage guidance.
49
questions
1
vote
0
answers
24
views
Global sensitivity of median queries
How to calculate the global sensitivity of median queries for the Laplace mechanism, if we know the lower and upper bound of the dataset?
- 11
0
votes
0
answers
14
views
distinguishing gradients of samples in semi supervised learning under mixup
I'm having trouble drawing conclusions from a specific semi supervised machine learning setting. My use case would be using mixup [Zhang2017] and label guessing similar to the mixmatch paper [...
- 1
0
votes
0
answers
65
views
$(\alpha, \beta, \epsilon)$- DP-PAC learner with sample size polynomial, what is meant with that?
In the case of a $(\alpha, \beta, \epsilon)$- DP-PAC learner (with finite hypothesis class $H_d$), what is meant by sample size polynomial in different expressions for d, $\alpha$, $\beta$, $\epsilon$?...
- 295
1
vote
1
answer
155
views
differential-privacy: show $\epsilon$ -differentially privacy
In this problem we consider a sensitive dataset $x \in \{−1, 1\}^n$. We consider the bounded setting where neighboring n-dimensional datasets differ in one coordinate. $A$ mechanism is available that ...
- 295
2
votes
0
answers
13
views
Differential Privacy guarantee that takes into account the approximate density (e.g., the pseudo randomness) used in practice?
In theory the differential privacy guarantee comes from adding randomness to an algorithm so whatever is output is a sample from a target distribution (e.g., the Laplacian, Gaussian, Exponential ...
- 411
1
vote
0
answers
17
views
How to quantify privacy changes for differential privacy with extra columns?
I just started learning about differential privacy (DP), and I've been trying to figure out how DP is affected when 1 or more column is added to the data. Specifically, suppose:
$$Pr(F(X) \in S) \leq \...
- 11
1
vote
0
answers
19
views
How does DP-SGD handle the composition problem between layers?
The paper 'Deep Learning with Differential Privacy' proposed DP-SGD and moments accoutant. The authors consider the NN layers separately, which allows they set different $C$ and $\sigma$ for different ...
- 11
1
vote
0
answers
69
views
Differential Privacy in Machine Learning: Why do we add noise to gradients and not the outputs?
I recently came across the concept of Differential Privacy in Machine Learning / Deep Learning. As far as I understand, usually, noise is added to gradients during the training epochs (also called ...
- 11
1
vote
0
answers
35
views
Verifying epsilon in differential privacy [closed]
Suppose I create a synthetic dataset out of a real dataset using some epsilon. How can one verify that my synthetic dataset is atleast epsilon differential private?
- 161
2
votes
0
answers
43
views
How should one calculate the confidence interval in 2020 census data, which uses differential privacy?
In 2020, the U.S. Census Bureau began injecting noise into census counts using a differential privacy technique. See here for a popular press description and here for some official literature.
The ...
- 131
1
vote
0
answers
66
views
Differential Privacy on tabular data
I wanted to create a deferentially private dataset of my private dataset. While I understand the concepts behind DP, I am unable to grasp how this can be applied to these fields. e.g.
...
- 161
1
vote
1
answer
29
views
Should data be anonymized irrespective of the domain? [closed]
I am using a dataset from Marketing and sales department. The dataset contains customer name (company name), company address, pincode, no of orders placed, revenue generated from that customer etc.
My ...
- 2,956
2
votes
2
answers
594
views
Why is DP-SGD differentially private?
The paper 'Deep Learning with Differential Privacy' explains how to make a deep learning algorithm as differentially private. This explanation is implemented in Tensorflow Privacy
My question is: we ...
- 21
2
votes
1
answer
67
views
adding noise to prediction task
Say that a teacher wishes to use a standard prediction task from Kaggle as a course assignment, and the idea is to have students submit their predictions, and award grades based on a test set (...
- 1,431
1
vote
1
answer
228
views
Why does Chebyshev's inequality yield that the probability of Laplacian noise being bigger than x is bounded like this?
I am trying to understand this proof of the bounds of Laplacian noise used in a paper on differential privacy.
Given a random variable $Lap\left ( \frac{\Delta f}{\varepsilon } \right )$, apparently ...
- 113
3
votes
0
answers
147
views
Adding Laplace noise to a learned neural network
My question is related to the concept of differential privacy and deep learning. I found many papers to learn neural networks with differential privacy, but is it also possible to achieve differential ...
- 31
1
vote
1
answer
44
views
Does $\epsilon$-differential privacy treat databases with one record of difference a completely different databases?
Does $\epsilon$-differential privacy treat databases with one record difference completely different database?
What I want to know is about continuous release.
Suppose we have a set of users and some ...
- 213
0
votes
1
answer
71
views
Can a folded LaPlace distribution (or other folded distributions) be used with Ɛ-differential privacy
I have a single value in (or over) our dataset, let's say a count of something, and we want to keep that value private within a certain range. This range is the sensitivity.
The adversary can ask if a ...
- 113
2
votes
1
answer
158
views
Upper Bound and Lower Bound on Means when Distributions are bounded?
Suppose we have two different probability distributions $p, q$ defined on input $x \in [0,1]$. We know that for any value of $x$ in the domain, we have $\exp^{-a} \leq \frac{p(x)}{q(x)} \leq \exp^{a} $...
0
votes
1
answer
73
views
Laplace Inequality
I am trying to prove that if $r_i \sim Lap(0,1/\varepsilon)$ where $\varepsilon >0$ then:
$$Pr[r_i \geq 1+r^*] \geq e^{-\varepsilon}Pr[r_i \geq r^{*}]$$.
I know that for $r*>0$ it satisfies ...
2
votes
0
answers
60
views
What is statistical difference?
In the paper "Calibrating Noise to Sensitivity in Private Data Analysis" by Dwork et al., the term "statistical difference" is used as following (in page 280):
Finally, if a $1 − \...
- 21
2
votes
1
answer
119
views
Laplace mechanism on vector record?
Does the definition of neighboring database in differential privacy capture the multi-dimensional record?
Let's say we have a database domain $\mathbb{N}^{n\times d}$ where $n$ is the number of ...
- 213
0
votes
1
answer
213
views
Lemma KL-Divergence (Differential Privacy)
I am studying differential privacy and I got stuck again in proof of a lemma. Which is:
"$D_{\infty}^\delta(Y||Z) \leq \epsilon$ if and only if there exists a random variable $Y'$ such that $\...
2
votes
0
answers
49
views
Interpretation of the ratio of two pdfs evaluated at a certain point?
What is the interpretation of the ratio of two pdfs evaluated at a certain point?
Is that a statistical distance?
Is there any applications of it?
I only know one application of differential privacy, ...
- 213
3
votes
1
answer
198
views
Does Laplace mechanism work only on linear query?
Does Laplace mechanism work only on non-multiplicative query?
For example, suppose a database (an array here) $\mathbf{x} = (x_1, \ldots, x_n)$.
Is it possible to do design laplacian mechanism for ...
- 213
2
votes
1
answer
33
views
Why $Pr[X-\mu \geq t]= Pr[e^{\lambda(X-\mu)} \geq e^{\lambda t}]$ for all $\lambda> 0$
I hope everyone is having a nice day. I don't know why this inequality holds.
$$
Pr[X-\mu \geq t]= Pr[e^{\lambda(X-\mu)} \geq e^{\lambda t}]
$$
For $\lambda >0$. I guess it has something to do ...
1
vote
0
answers
66
views
What data is sent from a client to a server in a federated learning setting?
So far, I thought federated learning works like this:
All clients have the same machine learning model (if not personalized). They have their unique data and then train this model (e.g., neural ...
- 11
2
votes
0
answers
59
views
Can we get the input from a multilayer perceptron based on the output of one of its hidden layers?
I was reading a relatively new paper that proposed to split a nerual networks layers into groups and sending each group to different nodes to train them in a distributed manner. In order to not send ...
- 121
2
votes
3
answers
1k
views
What is the difference between data perturbation and differential privacy?
I cannot distinguish the terms "data perturbation" and "differential privacy".
If the data perturbation is the process that adds some small value sampled from specific distributions such as Laplacian ...
- 213
1
vote
1
answer
2k
views
Global sensitivity of mean and variance in differential privacy?
Please explain me why global sensitivity of a mean or variance queries will be
(b-a)/n
and
(b-a)^2/n
where b is the upper ...
- 111
3
votes
1
answer
43
views
Communicating aggregate percentage changes in data without exposing individual contributors
So i have a dataset that tracks widget production from 100 different factories, each individually owned and highly competitive. Each line contains the factory name, the date of production, and the # ...
- 31
1
vote
0
answers
78
views
Is differential privacy applicable to summation value? [closed]
Given a collection of real data, we want to do some statistical analysis.
For example, we take sum of them and reveal it to someone else.
But, we want to guarantee some sort of privacy of ...
- 213
0
votes
1
answer
101
views
How can one apply differential privacy to this network dataset? [closed]
I've been reading up recently on differential privacy and I'm just starting to understand it. I've also read this paper that basically determined the sexual orientation of a user using Facebook ...
- 3
2
votes
2
answers
846
views
differentially private release of histograms (non-negative valued queries)
Two practical questions arise when releasing differentially private histograms/counts via addition of Laplace/Gaussian noise:
1) Is the result of noise addition truncated/rounded (since we know that ...
- 944
5
votes
1
answer
614
views
How to generate a sanitized dataset using Differential privacy?
I'm learning about differential privacy. I understand the concept behind differential privacy, that you can add a small noise to the query to mask the true value using transformations like Laplace or ...
- 243
5
votes
2
answers
3k
views
What are global sensitivity and local sensitivity in differential privacy?
I am learning differential privacy now, and there is no one surrounding I can ask questions about differential privacy. I am confused about the definitions of the global sensitivity and local ...
- 101
2
votes
1
answer
54
views
postprocessing additive noise in differentially private data
when releasing differentially private datasets we often have (or can plausibly assume) knowledge of the noise added to the data to achieve privacy - we can even have good approximations of the scale ...
- 944
2
votes
1
answer
131
views
Why does "sticky noise" defy averaging attack?
I have read an interesting paper (pdf) describing how a privacy preserving technique might be breached, but I am having trouble understanding the following paragraph describing one of several layers ...
- 607
0
votes
1
answer
312
views
How to test for differential privacy on multiple choice data?
I apologize I am new to statistics so I do not know all terms and concepts.
My current algorithm for adding noise to multiple-choice favorite color data is this:
...
- 1
5
votes
1
answer
260
views
Most powerful test bounds in differential privacy setting
I am interested in the setting of differential privacy- let's say a random function $\mathcal{D}:X\to\mathbb{R}$ discriminates between (distinct) $x, y \in X$ in a differentially private way if
$$
\...
- 2,155
2
votes
1
answer
196
views
Issue with the definition of differential privacy
The definition of differential privacy states that if $\mathcal{M}$ is $(\epsilon,\delta)$-differentially private, then $\forall x,y$ such that $||x-y||_1\leq1$ and for all $S \subseteq \mathrm{Range}...
- 23
5
votes
2
answers
105
views
Privacy through moving averages?
I am considering the following hypothetical situation: I have a time series of data. In general, 'the public' should have access to features of this data. However, making the time series available ...
- 221
4
votes
3
answers
330
views
Differential privacy of identity query
I am trying to understand some of the papers that present identity query mechanisms that satisfies differential privacy, for example the compressive mechanism which uses what they call a universal ...
- 41
3
votes
3
answers
305
views
Effect of exp(ϵ) in Differential Privacy Definition
I am reading about differential privacy and would like to understand the implications of the different values of $\varepsilon$ in the definition below:
$$\mathbb{P}[K(D_1) \in \mathcal{S}] \leqslant \...
- 31
3
votes
1
answer
242
views
Differential Privacy: why $\delta$ negligible on the row numbers?
The definition of differential privacy says that an algorithm $M$ is $(\epsilon,\delta)$-differentially private if
$$P(M(x \in D) \in S)\leq e^\epsilon P(M(x \in D')\in S) + \delta$$
where $D,D'$ ...
- 747
14
votes
2
answers
12k
views
What is meant by "Laplace noise"?
I am currently writing algorithm for differential privacy using the Laplace mechanism.
Unfortunately I have no background in statistics, therefore a lot of terms are unknown to me.
So now I'm ...
- 143
30
votes
4
answers
950
views
Has the journal Science endorsed the Garden of Forking Pathes Analyses?
The idea of adaptive data analysis is that you alter your plan for analyzing the data as you learn more about it. In the case of exploratory data analysis (EDA), this is generally a good idea (you are ...
- 20.2k
4
votes
2
answers
2k
views
What is the purpose of using a Laplacian distribution in adding noise for Differential Privacy?
I am reading up on Differential Privacy and it is mentioned that the technique relies on adding some controlled noise to the release of responses to queries towards a statistical database. This is ...
- 43
1
vote
0
answers
68
views
What current methods for statistical disclosure limitation are best trade-offs between data privacy and data utility?
In many situations raw microdata is not released by institutions due to privacy preserving. Many techniques are used for protecting sensitive values in data. But many of them can destroy multivariate ...
- 3,919