laplace noise mechanismanikolov/csc2412f20/slides/dp-lap.pdflaplace noise mechanism the laplace...
TRANSCRIPT
Laplace Noise Mechanism
Reminder: Counting queries
Given a predicate q : X ! {0, 1} (e.g., “smoker?”), we define the corresponding
(normalized) counting query
q(X ) =1
n
nX
i=1
q(xi ).
A workload Q of counting queries is given by predicates q1, . . . , qk .
Q(X ) =
0
B@q1(X )
...
qk(X )
1
CA 2 [0, 1]k .
E.g., “smoker?”, “smoker and over 30?”, “smoker and heart disease?”, etc.
9
" smoker,
heart disease,over 30 ? "
Answering counting queries with Randomized Response
10
How can I answer k counting queries w/ e . Dp
using RR ?
K instances of RR ( one per query )with (q) - Dp ) whole mechanismis E- DP by
composition than
All queries get answers we c- IL error
with prob . I - of
as long as n→kIe8l Exercise : Do the
£42 details !
Sensitivity
The `1 sensitivity of f : X n ! Rk is
�1f = maxX⇠X 0
kf (X )� f (X 0)k1 = maxX⇠X 0
kX
i=1
|f (X )i � f (X 0)i |
Measure of how much a person can influence f .
11
e.g. textQQ - f }
% neighbouringe.g , if f :D
"→ IR
then ns.f-zinqylfkl- f 't'll
Sensitivity of a workload of counting queries
12
Suppose flx ) -
- alt) -- (g! It, ) for a workload
of k counting queries
Give an upper bound on D,Q
gilt) - t . gil qilxj) c- soil 's D, q .
.= In ki
1a --
em. Ei !9i"?ia%"" ' a.
Laplace noise mechanism
The Laplace noise mechanism MLap (for a function f : X n ! Rk) outputs
MLap(X ) = f (X ) + Z ,
where Z 2 Rk is sampled from Lap(0, �1f" ).
Lap(µ, b) is the Laplace distribution on Rk with expectation µ 2 Rk and scale b > 0,
and has pdf
p(z) =1
(2b)ke�kz�µk1/b =
1
(2b)kexp
�1
b
kX
i=1
|zi � µi |!
13
Z. .
. .. .
Z,are independent
Zi is from a one - dimensional Laplo,
µwpise-DpTy
Privacy of the Laplace noise mechanism
For any f , MLap is "-DP.
Let X ⇠ X 0, and let p be the pdf of M(X ), and p0 the pdf of M(X 0).
p(z) =
✓1
2�1f
◆k
e�"kz�f (X )k1/�1f p0(z) =
✓1
2�1f
◆k
e�"kz�f (X 0)k1/�1f
Claim: enough to show maxz2Rkp(z)p0(z) e"
14
THE;fCXnhapHH.FI/=flHltZLap lap
~ Laplftx't . At)E E
Henk → Ethel'IPCMLAPIXIES ) - 1PM dZ Else' p' czidz
= ee fgptztdz = e' IPCMCX
' ) c- S ).
Privacy of the Laplace noise mechanism
p(z) =
✓1
2�1f
◆k
e�"kz�f (X )k1/�1f p0(z) =
✓1
2�1f
◆k
e�"kz�f (X 0)k1/�1f
15
Need ;Kzn ¥7, ee
'
Y E
FI= exp (Ff ( UZ - TH
'
) ",- Hz- HNK
, )) e eep'
IZ) my
←EHFIX ) - fit 'm ,
HZ - fl X' Ill , E HZ - fl till, t k fit ) - fix 'll! Triangle inequality
¥i
Histograms
Suppose the query workload Q = {q1, . . . , qk} “partitions” the data.
• 8x 2 X : at most one of q1(x), . . . , qk(x) equals 1.
• e.g., “votes for the Liberal Party per riding”
What is the sensitivity?
16
"' mi.
a
X of 4 . . ..
.Xu 's
- F '= 94 . .
. .
. ti'
,. . . .tn }
fray,"QIN - QH 'll! -
-
finger, ÷! Iq .
-
K ) - gilt 'll E E
because E 2 queryvalues
change by c- In each.
Accuracy of the Laplace noise mechanism
If Z 2 Rk is a Laplace random variable from Lap(µ, b), then, for every i
P(|Zi � µi | � t) = e�t/b.
17
Generalizes to " K - norm mechanism "
Hardt,Talwar
-
Map (H - Lap ( fit ) , diet )Rl ?at llk.pk/i-fCXIilza1z?zlPllkaplxt.-fctHzH
E k . e- debit
,
then is ,ft kn,
Minar error z d) else-1¥
E.g . if f- (x) -- Q (H E B
i.e.,answers to
k counting if n zHulkingqueries he