READ IT!

1. PRIVACY-ASSURED OUTSOURCING OF IMAGE
RECONSTRUCTION SERVICES IN THE CLOUD
PRESENTED BY : UNDER GUIDENCE OF :
THAHIRA A RAJI R PILLAI
S7 CSE ASSISTANT PROFESSOR
ROLL NO : 61 DEPARTMENT OF CSE

2. OVERVIEW
 INTRODUCTION
 RELATED WORK
 PROBLEM STATEMENT
 OIRS DESIGN
 EMPIRICAL EVALUATION
 FUTURE SCOPE
 CONCLUSION
 REFERENCES
2
1/13/2015

3.  Today, there is a fast growing trend to outsource the image management
system to the cloud for its abundant computing resources and benefits.
How to protect the sensitive data while enabling outsourced image
services ?.........
Outsourced Image Recovery Service (OIRS) architecture.
INTRODUCTION
3
1/13/2015

4. OIRS
• addresses the design challenges of security ,
complexity , and efficiency simultaneously.
OIRS
• not only supports the typical sparse data service
but can be extended to non sparse general data
1/13/2015
4
Security Effectiveness
Efficiency Extensibility
DESIGN GOALS

5. LITERATURE SURVEY
1/13/2015
5
Sl
no
Title Strength Open end
1 “Compressed sensing”-
D.Donoho
Compressed sensing Basic properties
only
2 “Compact storage of correlated data for
content based retrieval”-
A.Divekar and O.Erosy
Leverage the compressed sensing
to compress the storage of
correlated image datasets
Does not consider
security
3 “The secrecy of compressed sensing
measurements”-
Y.Rachlin and D.Baronm
Explore the inherent security
strength of linear measurement
provided by compressed sensing
Not suited for all the
conditions

6. A. SERVICE MODEL AND THREAT MODEL
PROBLEM STATEMENT
1/13/2015
6

7. 1/13/2015
7
FIGURE2. Empirical results on the effectiveness of OIRS ,(a)-(a3) Original image ,(b)-(b3)
Reconstruction via encrypted data.(c)-(c3) Reconstruction via decrypted data

8. B. PRELIMINARIES
• Consider an n× 1 sparse data x. After sampling ,get an m×1 sample
vector,
y=Rx
R-m ×n selecting matrix
•The real world data x might not always be sparse.
•But it can be represented as a sparse vector f,(f Є ) under some properly
chosen Orthonormal basis V via x = Vf.
So,
y= Rx = RVf = Af where A = RV
1/13/2015
8
Compressed Sensing

9. •Plays an important role in the framework of compressed sensing
•Image reconstruction is equivalent to solve a LP problem as below,
min . r
subject to y = Af , -r <= f <= r
where, r- n×1 vector with positive real variables
•To make a standard form, replace f and r by
f = u-v
r = u+v
• Denote g=[u ,v] Є and F=|A,-A| Є 1/13/2015
9Linear Programming
𝟏 𝑻
min 1 𝑇 . g
Subject to y = F.g , g>=0
LP denoted as ,
Ω = ( F,y,I,1 𝑇 )

10. OIRS DESIGN
A.FRAME WORK AND SECURITY DEFINITION OF OIRS
The design challenge in OIRS is how the cloud efficiently solve optimization problem , Ω =(F,y,I, 1 𝑇) for
image reconstruction
The proposed OIRS meet the design challenges through random transformation based framework , which
includes four probabilistic polynomial time algorithms.
The framework of OIRS can be denoted as ,
= (KeyGen,ProbTran,ProbSolv,DataRec)
1/13/2015
10
KeyGen:
• Generates
the secret
key K
ProbTran:
• Generates a
randomly
transformed
Optimization
Problem Ωk
ProbSolv :
• Solves
transformed
problem Ωk &
Generates answer h
DataRec :
• Generates
answer g of
Ω

11. B. THE BLUEPRINT OF THE PROBLEM
TRANSFORMATION
 To make the algorithm ProbSolv to be a standard LP solver
,they use a series of random linear transformation steps over
objective function ,constraints , and feasible region of original
problem Ω
1/13/2015
11

12. 1.Use a random generalized permutation matrix π with positive entries.
2.Randomly pick an 2n×2n invertible matrix Q,and a 2n×1
vector e to protect the solution g via affine mapping g=Qh-e
3. Multiply a random 2n ×m matrix M to equality constraints and
later mix the result together with the inequality constraint
4. Multiply a random m ×m invertible matrix P to the both sides
of equality constraints
1/13/2015
12
Transformation procedure
min 1 𝑇 . 𝑔
Subject to y = F. g,
π.g>=0
min 1 𝑇 .(Qh - e)
subject to F . Q . h = y + F . e,
π . Q . h >= π . e
min 1 𝑇
. (Qh- e)
subject to F . Q . h = y + Fe,
(π - MF)Qh >= πe - M( y +Fe)
min 1 𝑇 .(Qh-e)
subject to PFQ . h = P . (y + Fe)
(π- MF)Qh>= πe-M(Y+Fe)
min 1 𝑇 . g
Subject to y = F.g ,
g>=0
min 1 𝑇
.(Qh-e)
subject to PFQ . h = P . (y + Fe)
(π- MF)Qh>= πe-
M(Y+Fe

13.  To make the randomly transformed problem sharing the same structure as Ω
1. make 1 𝑇
. Q is equal to 1 𝑇
2.make right hand side of the inequality constrains ,
r’ = πe- M(y + F e), always zero just as Ω
If ignore the constant term 1 𝑇
. e in objective function,then
the random LP,
where F’ = PFQ , y’ = P . (y + F . e) and π’=(π.Q-MFQ)
So the problem can be denoted as,
Ωk = (F’,y’,π’, 1 𝑇 ) 1/13/2015
13
min 1 𝑇 . h
subject to y’=F’ . h, π‘h>=0,

14. C . THE SCHEME DETAILS
 Two reasonable assumption about the informaton transformed
between the data owner and users,
1.a master secret key sk is used to generate random
sampling matrix R and secret key K for each image.
2.an orthonormal basis V , with which the image data x
can be represented as a sparse vector f,
1/13/2015
14

15. 1/13/2015
15

16. 1/13/2015
16

17. 1/13/2015
17

18. 1/13/2015
18

19. a, Data owner computes σ <- F(sk,s)
.He then uses σ as coins to sample
R and generates a secret key K =
(P,Q,e,π,M) from KeyGen( 1 𝑇
,σ).
b, He acquires the sample y .With F
= [RV. -RV] and y, he calls
ProbTran1(K,(y,F)) to encrypt y as
y’ and sends (y’,s) to cloud.
1/13/2015
191 . DATA SAMPLING PHASE
a, Data owner computes σ <-
F(sk,s) .He then uses σ as coins to
sample R and generates a secret
key K = (P,Q,e.π.M) from
KeyGen(1 𝑇
,σ).He calls
ProbTran2(K,F) to get (F’,π) and
sends to cloud
b,With Ωk ,the cloud calls
ProbSolv(Ωk) to output answer
h to user ,together with seed s.
c,The user computes σ <- F(sk,s)
,and uses σ to generate the key K
from KeyGen(1 𝑇).He then calls
DataRec(K,h) to get g =Qh-e and
recovers the image x=Vf, where f is
derived from g
2 . IMAGE RECOVERY PHASE

20. FUTURE SCOPE & ONGOING WORKS
1/13/2015
20
SPEEDUP WITH HARDWARE BUILT-IN DESIGN
• Hardware built-in design with great benefits in achieving the
secure OIRS with best possible service performance and user
experience.

21. A. EXPERIMENT SETTING
 The data owner/user and the cloud side process is implemented in MATLAB and use
the MOSEK optimization toolbox(http://www.mosek.com) as the LP solver
B. EFFICIENCY EVALUATION
To measure the efficiency of the proposed OIRS , Specifically to focus on the
computational cost of privacy assurance done by the data owner and data users ie, local side
and the cost done by the cloud side
EMPIRICAL EVALUATION
1/13/2015
21

22. TABLE1. Preleminary efficiency evaluation results of OIRS.Here
denotes the original image recovery time , the transformation time by data
owner ,and the decryption time by data user,respectively
1/13/2015
22

23. To evaluate how much computational savings OIRS can provide
to data owner/user , calculate a variable,
From the table we can see that ,OIRS can bring more than 3.4×
savings for the selected size image blocks
1/13/2015
23
assymmetric speedup=asymmetric speedup =

24. C. EFFECTIVENESS EVALUATION
1.CORRECTNESS EVALUATION
For correctness of the design ,all the images after
transformation and later recovered on the data user side ,
still preserves the same level of visual quality as the original
images.
1/13/2015
24

25. Reconstructed image quality increases along with the number
of measurements and the more the better
FIGURE3 . Comparison of recovered images using different number of
measurements m in OIRS.(a)m=128,(b) m= 192,(c) m=256.
1/13/2015
25

26.  OIRS,an outsoursed image recovery service from compressed sensing with
privacy assurance
With OIRS, Data owners can utilize the benefit of compressed
sensing
 Data users can leverage cloud’s abundant resources
CONCLUSION
1/13/2015
26

27. 1/13/2015
27
ADVANTAGE
APPLICATIONS
• Simple and Efficient
• Robustness and effectiveness in
handling image reconstruction
• MRI in health care system
• Remote sensing in geographical
system
• Military image sensing in mission
critical context

28. M. Atallah and K. Frikken, ``Securely outsourcing linear algebra computations,''in
Proc. 5th ASIACCS, 2010, pp. 4859.
 E. Candès and M. Wakin, ``An introduction to compressive sampling,''IEEE Signal
Proc. Mag., vol. 25, no. 2, pp. 2130, Mar. 2008.
 A. Yao, ``Protocols for secure computations (extended abstract),'' in Proc. FOCS,
1982, pp. 160164.
REFERENCES
1/13/2015
28

29. 1/13/2015
29