3. Introduction
Literature review
Contributions
Models and assumptions
Technique preliminaries
Proposed scheme
Performance evaluation
Conclusion
References
3
CONTENTS

4. Neural networks.
Back-propogation.
Improves the accuracy.
Joint/Collaborative learning.
4
INTRODUCTION

5. Challenges:
 To protect each participant’s private data set and
intermediate results.
 The computation/ communication cost introduced to
each participant shall be affordable.
 For collaborative training, training data is arbitrarily
partitioned.
5
INTRODUCTION(Contd..)

6. Provides privacy preservation for multiparty .
Collaborative BPN network learning over arbitrarily
partitioned data.
Guarantees privacy and efficiency.
Support multiparty secure scalar product.
Allow decryption of arbitrary large messages.
6
CONTRIBUTONS

7.  System Model:
 Trusted authority.
 The participating parties ( data owner).
 The cloud servers ( or cloud).
 Security Model:
7
MODELS AND ASSUMPTIONS

8.  Arbitrarily Partitioned Data
 Z parties (Z > 2 ) : Ps , 1 ≤ s ≤ Z.
 Database D with N rows : {DB1,DB2, ….. DBN}.
 Each row DBv ,1 ≤ v ≤ N has m attributes {xv
1 , xv
2 , xv
3 …..
xv
m}.
 DBv = DBv
1 U DBv
2 U DBv
3 U ….. U DBv
z .
 Each DBv, Ps has ts
v attributes.
8
TECHNIQUE PRELIMINARIES

9.  BACK –PROPOGATION NEURAL NETWORK LEARNING
9
TECHNIQUE PRELIMINARIES(Contd..)

10.  BGN Homomorphic Encryption
 Operations on plaintexts to be performed on their
respective cipher texts.
 Public-key “doubly homomorphic” encryption
scheme(called “BGN” for short).
 One multiplication and unlimited number of additions.
 Given ciphertexts C(m1) , C(m2) and C(m^1), C(m^2 ), one
can compute C(m1 m^1 + m2m^2) without knowing the
plaintext.
10
TECHNIQUE PRELIMINARIES(Contd..)

11.  PROBLEM STATEMENT
 3 layer (a-b-c configuration) neural network .
 N samples for learning data set .
 Arbitrary partitioned into Z( Z≥2) subsets.
 SCHEME OVERVIEW
 Each party encrypt her/his input data set.
 Participants upload the encrypted data to cloud.
 Cloud servers perform the operations.
 Secret sharing algorithm.
11
PROPOSED SCHEME

12.  PRIVACY PRESERVING MULTIPARTY NEURAL
NETWORK LEARNING
12
PROPOSED SCHEME(Contd..)

13. 13
PROPOSED SCHEME(Contd..)

14.  SECURE SCALAR PRODUCTION AND ADDITION WITH
CLOUD
Algorithm 3: Secure Scalar Product and Addition
 Key Generation.
 Encryption.
 Secure Scalar Product.
 Secure Addition.
 Decryption.
14
PROPOSED SCHEME(Contd..)

15.  SECURE SHARING OF SCALAR PRODUCT AND SUM
15
PROPOSED SCHEME(Contd..)

16.  APPROXIMATION OF ACTIVATION FUNCTION
16
PROPOSED SCHEME(Contd..)

17.  Experimental Evaluation
 Experiment Setup
• Amazon EC2 cloud.
• 10 nodes with 8-core 2.93-GHz Intel Xeon CPU.
• 8-GB memory.
• Testing data sets(Iris,kr-vs-kp,diabetes).
17
PERFORMANCE EVALUATION

18.  EXPERIMENTAL RESULT
18
PERFORMANCE EVALUATION(Contd..)

19.  EXPERIMENTAL RESULT (Contd..)
19
PERFORMANCE EVALUATION(Contd..)

20.  EXPERIMENTAL RESULT (Contd..)
20
PERFORMANCE EVALUATION(Contd..)

21.  EXPERIMENTAL RESULT (Contd..)
21
PERFORMANCE EVALUATION(Contd..)

22.  EXPERIMENTAL RESULT (Contd..)
22
PERFORMANCE EVALUATION(Contd..)

23.  EXPERIMENTAL RESULT (Contd..)
23
PERFORMANCE EVALUATION(Contd..)

24.  EXPERIMENTAL RESULT (Contd..)
24
PERFORMANCE EVALUATION(Contd..)

25.  ACCURACY ANALYSIS
 Accuracy loss in approximation of activation function.
 Maclaurin series used – accuracy can be adjusted by
modifying number of series terms.
25
PERFORMANCE EVALUATION(Contd..)

26. Secure and practical multiparty BPN network learning.
Cost independent of number of parties.
Scalable efficient and secure.
26
CONCLUSION

27. 1) N. Schlitter A Protocol for Privacy Preserving Neural
Network Learning on Horizontal Partitioned Data, Proc.
Privacy Statistics in Databases (PSD ’08), Sept. 2008
2) T. Chen and S. Zhong, Privacy-Preserving
Backpropagation Neural Network Learning,IEEE Trans.
Neural Network, vol. 20, no. 10, Oct. 2000,pp. 1554-1564
3) A. Bansal, T. Chen, and S. Zhong, Privacy Preserving
Back-Propagation Neural Network Learning over
Arbitrarily Parti-tioned Data,Neural Computing
Applications,vol. 20, no. 1, Feb. 2011, pp. 143-150,
4) D. Boneh, E.-J. Goh, and K. Nissim, Evaluating 2-DNF
Formulas on Ciphertexts,Proc. Second Int’l Conf. Theory
of Cryptography (TCC ’05), pp. 325-341, 2005.
27
REFERENCES

28. THANK YOU