投影片 1

appliancepartAI and Robotics

Oct 19, 2013 (4 years and 19 days ago)

86 views

Privacy
-
Preserving Self
-
Organizing Map

Shuguo Han and Wee Keong Ng


Center for Advanced Information Systems, School of Computer
Engineering,Nanyang Technological University, Singapore

(DaWak 2007)

1

2009/11/02

Outline


Introduction


SOM


Privacy
-
preserving SOM protocol


Conclusion


2

2009/11/02

Introduction


various data mining algorithms have been enhanced
with a privacy preserving version for horizontally
and/or vertically partitioned data



propose a protocol for privacy
-
preserving self
-
organizing map for vertically partitioned data
involving two parties.

2009/11/02

3

SOM


Self
-
organizing map (SOM) is
awidely

used
algorithmfor

transforming data sets to a lower dimensional space to facilitate
visualization


To projection of the data set while preserving the topological

properties of the data set.


SOM is a feed
-
forward neural network without any hidden layer
adjusting input.

2009/11/02

4

SOM



Competition phase,
不斷學習使輸出與目標值能達到相同值後結束


input data


X(
t) = [Xi(t),X2(t), . . . ,
Xd
(t)]


each neuron’s weight vector (randomly
initial weight )


W
j
(t) =
[W
j,1(t),Wj,2(t), . . . ,
Wj,d
(t) ]



2009/11/02

5

SOM


Euclidean distance:





Winner neuron:



Update weight vector


2009/11/02

6

Z
為學習函式
,
越大表學習越快
,
一般介於
ま1
之間

Privacy
-
preserving SOM protocol


Protocol 1. Privacy
-
Preserving Self
-
Organizing Map




,
the weight vector
holds two private component vector. At step
t=0 where and

are securely and
randomly generated respectively



2009/11/02

7

Privacy
-
preserving SOM protocol


input data X = (X1,X2, . . . , Xd) from
feature space















The different between SOM and stand SOM is that the
subprotocol are required to perform some computations
securely

2009/11/02

8

Privacy
-
preserving SOM protocol


Protocol 2. Secure Computation of
Closest Cluster Protocol


=> to find winner neuron



by
applying
the secure scalar product protocol [2,4].




2009/11/02

9


2009/11/02

10

Privacy
-
preserving SOM protocol


Correctness

2009/11/02

11

Privacy
-
preserving SOM protocol


Protocol 3. Secure Weight Vector
Update Protocol

2009/11/02

12

// adjust all neuron’s weight vector

// j is how many neurons in this grid

// i is how many attributes of input

Privacy
-
preserving SOM protocol


Correctness


2009/11/02

13

Privacy
-
preserving SOM protocol


Protocol 4. Secure Detection of
Termination Protocol

2009/11/02

14

Privacy
-
preserving SOM protocol


Correctness










stop

2009/11/02

15

Conclusion


(1) to securely discover the winner
neuron from data privately held by two
parties


(2) to securely update weight vectors of
neurons


(3) to securely determine the
termination status of SOM.

2009/11/02

16