Lecture6 0

Information Security – Theory vs.
Reality
0368-4474, Winter 2013-2014
Lecture 6:
Machine Learning Approach
to Side-Channel Attacks
Guest lecturer: Guy Wolf
1
Machine Learning Approach
to
Side-Channel Attacks
Guy Wolf
Tel Aviv University
guy.wolf@cs.tau.ac.il
2013
Guy Wolf (TAU) Machine Learning & Side-Channels 2013 1 / 61

Outline
1 Introduction
Side-Channel Leaks
Machine Learning
2 Quick Recap: Correlation Power Analysis
3 Alternative Attack: Template Power Analysis
Dimensionality Reduction
Classification
Power Trace Alignment
4 Information & Activity Leaks
Hidden Markov Model
Acoustic Analysis
5 Conclusion

Side-Channel Leaks

Side-Channel Leaks
Power Consumption
Y
H
HH
H
H

Side-Channel Leaks
Probing
Power Consumption 6
Y
H
HH
H
H

Side-Channel Leaks
Probing
Power Consumption 6
Y
H
Response Times
HH *

H
H


Side-Channel Leaks
Probing
Power Consumption 6
Y
H
Response Times
HH *

H
H

HH
HH
j
H
Acoustics

Side-Channel Leaks
Probing
Power Consumption 6
Y
H
Response Times
HH *

H
H

HH
HH
j
H
Acoustics
?
Electromagnetic Radiation

Side-Channel Leaks
Probing
Power Consumption 6
Y
H
Response Times
HH *

H
H

HH

HH
Temperatures j
H
Acoustics
?
Electromagnetic Radiation

Processing Leaked Data

leaked data
-

leaked data
-

leaked data
-
-

leaked data
-
-
-

leaked data
-
+)
⊆ RO(100
-
-
-

leaked data
-
+)
⊆ RO(100
-
-
-



Machine Learning
Machine learning encompasses tools that perform smart analysis of

data, such as:
Discovery of useful, possibly unexpected, patterns in data
Non-trivial extraction of implicit, previously unknown and
potentially useful information from data
Exploration & analysis, by automatic or semi-automatic means,
of large quantities of data in order to discover meaningful
patterns
Common tasks include:
Dimensionality reduction
Clustering & Classification
Regression & Out-of-sample extension

Quick Recap:
Correlation Power Analysis

Quick Recap: Correlation Power Analysis
Key k -
- Ciphertext
OC
- C
Plaintext p C
C
C
C
Internal value y = f (k, p)
Typical assumption:
power consumption correlates with Hamming Weight of y

Plaintexts: Traces: Internal Vals.: Hamm. Wts.:
p1 - - 00101001 - 3
.. .. .. ..
. . . .
pn - - 10011011 - 5

Traces matrix: Hamming Weight matrix:
times key candidates
:::: trace x :::: 1
.. .. .. .. .. .. ..
plaintexts
plaintexts
. ......
:::: trace x :::: n


~
:::: trace x ::::
P 1
P PP
..
Hamming weight of
.. .. P
.. j
.. ..
plaintexts
plaintexts
internal value for
.
designated plaintext
. . . ..

and key candidate

:::: trace x :::: n

:::: trace x :::: 1
.. .. .. .. .. .. ..
plaintexts
plaintexts
. ......
:::: trace x :::: n
Correlation indicates the correct time and key

:::: trace x :::: 1
Pearson Coefficient:
cov(x ,y )
linear correlation std(x
.. .. .. .. .. .. ..
plaintexts
plaintexts
)std(y )
Mutual Information:
.
general statistical correlation
(Entropy(x ) + Entropy(y ) − Entropy(x , y ))
......
:::: trace x :::: n
Correlation indicates the correct time and key

Alternative Attack:
Template Power Analysis

Alternative Attack: Template Power Analysis
Training: Learn traces of many plaintexts & keys
times
:::: trace ::::
:::: trace ::::
:::: trace ::::
:::: trace ::::
plaintexts & keys
:::: trace ::::

:::: trace ::::
:::: trace ::::
..
.
times
:::: trace ::::

*



plaintext = p1 :::: trace :::: key = k0

:::: trace ::::
A @ 

A @
A @
A
R
@ :::: trace :::: 


:::: trace ::::

A
A
key = k1
:::: trace ::::


A 
A
U
A
:::: trace ::::
..
.
times
:::: trace ::::

*



plaintext = p1 :::: trace :::: key = k0

:::: trace ::::
A @ 

A @
A @
A
R
@ :::: trace ::::  

:::: trace ::::  key = k1

A
A
A :::: trace :::: s in gle t r a c

e
A
usin g a
U
A
::::
Online:
Attack trace ::::
..
.
Representing power traces as vectors
Essentially, power traces are high-dimensional vectors with

m = O(100+) of measurements.
:::: trace x ::::

Rm

x[1] , x[2] , ··· , x[m] ∈
Traces can be analyzed as vectors in the Euclidean space m with `2 R

norms & distances
Somewhat arbitrary representation, but effective and convenient

Learning & analyzing trace vectors
Template assumption: the positions of traces in Rm (with `2 norm)

correlate with the plaintext & key values that generated them
(p0 , k0 )
(p1 , k1 )
(p2 , k2 ) (p3 , k3 )
Correlation is either directly seen or via some internal value

Ideally, correlation is expressed as clustering by plaintext & key
Theoretically, machine learning tools can be used to detect clustering
Example: classic TPA attack using Gaussian statistical modeling1
Model power consumption of encrypting plaintext p with key k as a

~ (k,p) ∼ N (~
random variable X µ(k,p) , Σ(k,p) )
~ (k,p) is drawn from a multidimensional normal (Gaussian) distribution
X
~ (k,p) ∈
µ Rm is the mean power consumption for k & p
Σ(k,p) is the m × m noise covariance matrix for k & p
The likelihood of a trace ~x originating
from k & p is
−1/2
m
L(k,p) (~x ) = (2π) |Σ(k,p) | ~ (k,p) )T Σ−1
exp − 12 (~x − µ (k,p) (~
x − µ
~ (k,p) )
Attack single trace (with known plaintext p):

Compute likelihood of it originating from each key candidate (with p)
Choose key candidate with maximum likelihood
1
“Template Attacks”, 2003, by S. Chari, J.R. Rao, and P. Rohatgi
Curse of dimensionality
Unfortunately, directly analyzing high-dimensional vectors is usually

unfeasible due to the “Curse of Dimensionality”.
Curse of Dimensionality
A general term for various phenomena that arise when
analyzing/organizing high-dimensional data.
Common theme - difficult/impractical/impossible to obtain statistical
significance due to sparsity of the data in high-dimensions
Causes poor performance/results of statistical methods compared to
low-dimensional data
Common solution - use dimensionality reduction methods and analyze their

resulting embedded space.
Example: only use the ` < m time indices that provide the highest differences
between mean power consumptions of different key-plaintext pairs.
with
Principal Component Analysis

Principal Component Analysis (PCA)
3D space
6
-


3D space
6
-


Assume:
avg = 0
3D space
6
-

Find:
max variance directions

Principal Component Analysis (PCA) - covariance matrix
traces
times























times
Covariance

=


matrix


traces










tracei [t1 ] · tracei [t2 ]

P
cov (t1 , t2 ) ,
i

Principal Component Analysis (PCA) - spectral theorem
Eigenvector
H
Hj
H
t1
Spectral theorem Eigenvalue
@
applies to cov. R
@
λi φi = Covariance
matrix φi
matrices:
t2
Eigenvectors
H
HH
H

j
H
@
@
Covariance @
matrix = @
@ SVD (Singular Value
@
@
@
@ Decomposition)
6
Eigenvalues
P
Spectral Theorem: cov (t1 , t2 ) = λi φi (t1 ) φi (t2 )
i
Principal Component Analysis (PCA) - truncated SVD
eig
en
va
lu
es
λ1 λ2 λ3 λ4 λ5
Many datasets (incl. power traces) have a decaying cov. spectrum

Principal Component Analysis (PCA) - truncated SVD
Eigenvectors
H
HH
H

j
H
@
@
Covariance @
matrix = @
@
@
@
@@
| {z }
principal components 6
Eigenvalues
Approximate cov. matrix by truncating small eigenvalues from SVD

Principal Component Analysis (PCA) - example
Consider simple case of traces that are all on the same high
dimensional line
~
Straight line is defined by a unit vector ψ =1
Points on the line are defined by multiplying ψ ~ by scalars

The traces can be formulated as xi = ci ψ ~
P P ~ ~
Covariance: cov(t1 , t2 ) = xi [t1 ]xi [t2 ] = ci ψ[t1 ]ci ψ[t2 ] =
i i
~
( c 2 )ψ[t
P ~ ~ 1 )ψ(t
c k2 ψ(t
1 ]ψ[t2 ] = k~
~ 2) ~c , (c1 , c2 , . . .)
i
i

dimensional line
~
Points on the line are definedkck

by2 multiplyingψ ψ ~ by scalars
Covariance
Covariance: = P
ψ P ~
cov(t1 , t2 ) = xi [t1 ]xi [t2 ] = ci ψ[t
Matrix i
~
1 ]ci ψ[t2 ] =
i
~
( c 2 )ψ[t
P ~ ~ 1 )ψ(t
c k2 ψ(t
1 ]ψ[t2 ] = k~
~ 2) ~c , (c1 , c2 , . . .)
i
i

dimensional line
~
Points on the line are defined by multiplying ψ ~ by scalars

P P ~ ~
Covariance: cov(t1 , t2 ) = xi [t1 ]xi [t2 ] = ci ψ[t1 ]ci ψ[t2 ] =
i i
~
( c 2 )ψ[t
P ~ ~ 1 )ψ(t
c k2 ψ(t
1 ]ψ[t2 ] = k~
~ 2) ~c , (c1 , c2 , . . .)
i
i
Covariance matrix has a single eigenvalue k~c k2 and a single

~ which defines principal direction of the trace-vectors
eigenvector ψ,

f
v
f
v
f
v
3D space
6 f
v
-
f
v


f
v
f
v
φ1 = ψ
f
v

3D space
6 f
v
-
f
v


3D space
6
-


Length: eigenvalues
λ1 φ1
Direction: eigenvectors
λ2 φ2
3D space
I
@
6
-

principal components ⇒ max var directions

Principal Component Analysis (PCA) - projection
Projection on principal components:

Principal components




















Traces














Principal Component Analysis (PCA) - projection
Projection on principal components:
3D space λ1 φ1
6

-
7→
1D space
-

PCA algorithm
PCA algorithm:
1 Centering
2 Covariance
3 Eigendecomposition
4 Projection
Alternative method: Multi-Dimensional Scaling (MDS) - preserve

distances/inner-products with minimal set of coordinates.
Short tutorial on PCA & MDS:

www.cs.haifa.ac.il/∼rita/uml course/lectures/PCA MDS.pdf

Summary
Traces (high-dim. vectors): Projected traces

(low-dim. vectors):
.. 7→ PCA

Summary
Traces (high-dim. vectors): Projected traces

(low-dim. vectors):
.. 7→ PCA
.
Next task: how to find keys from the low-dimensional vectors?
Clustering & Classification
with
Support Vector Machine

Classification
Clustering
Cluster analysis
Clustering - the task of grouping
objects such that objects in the
same cluster are more similar to
each other than to those in other
clusters.

Classification
Clustering
Cluster analysis
Clustering - the task of grouping
objects such that objects in the
same cluster are more similar to
each other than to those in other
clusters.
Learning types
Unsupervised learning: Supervised learning:
Trying to find hidden Inferring functions from
structures in unlabeled data. labeled training data.

Classification
Clustering & classification approaches
Classic TPA using Gaussian statistical models:

The analysis considers many clusters (one for each key-plaintext pair)
Clusters are assumed to correlate with normally distributed random
~ (k,p) ∼ N (~
variables X µ(k,p) , Σ(k,p) )
Requires many traces for each key-plaintext pair to compute µ
~ (k,p) & Σ(k,p)
Simplified bit clustering with Support Vector Machine (SVM):

Classify each bit separately - only two classes are considered for each bit
Requires less training traces than classic TPA - traces are grouped by
bit values, not by the key value
No statistical assumptions required - geometric classification using a
separating hyperplane

Classification
Simplified bit clustering - only two classes
bit i of key
7→ 0 B 1
BN
(p1 , k1 )
(p0 , k0 )
(p2 , k2 ) (p3 , k3 )

Classification
Support Vector Machine (SVM) - separation with hyperplane

Classification
b=0
b=1

Classification

b=0

b=1
Guy Wolf (TAU)

Machine Learning & Side-Channels 2013 30 / 61
Classification

b=0

b=1
Guy Wolf (TAU)

Machine Learning & Side-Channels 2013 30 / 61
Classification
b=0
b=1

Classification

b=0

b=1

Classification
Support Vector Machine (SVM) - quantifying robustness with margins

b=0

b=1

Classification
Support Vector Machine (SVM) - quantifying robustness with margins

b=0

b=1

Classification
SVM formulation - hyperplane

~ · ~x > 0
w

~
w
Y0
H
~ · ~x < 0
w

Classification
SVM formulation - hyperplane with margin

~ · ~x ≥ α
w




~
w


αH
Y0 
~ · ~x ≤ −α
w

Classification
SVM formulation - shifted hyperplane with margin
~ · ~u
c=w

~ · ~x − c ≥ α
w




~
w

Y~u 
αH

~ · ~x − c ≤ −α
w

Classification
SVM algorithm
SVM training
Input:
Points {~xi } from PCA of the traces
Labels {bi } according to attacked bit:

1 bit is 0
bi =
−1 bit is 1
Solve the quadratic program (e.g., using Lagrange multipliers):
Find max α
~ · ~xi − c ≥ bi α
s.t. w
Output: the solution (~

w , c, α)

Classification
SVM algorithm
SVM classifier
Input:
New point ~x from PCA projection of attacked trace
The solution (~
w , c, α) from SVM training
~ · ~x − c:
Classify by value of w
attacked bit is 0 attacked bit is probably 1

z }| { z }| {
α
−α | {z } | {z }
attacked bit is probably 0 attacked bit is 1

Power Analysis with PCA & SVM
based on recent paper2
SVM - 1st bit

..
- PCA .
@@
R SVM - 7th bit
Use single PCA & multiple SVMs (one per bit)

to learn traces (in training phase) and attack a key byte
2
“Side channel attack: an approach based on machine learning”, 2011,
by L. Lerman, G. Bontempi, and O. Markowitch
Power Analysis with PCA & SVM
PCA results (colored by specified bit values):
Effective SVM attack: SVM attack will fail:
Empirical results (on 3DES): desc. success rate (by bit position)
7th bit success rate: ∼95% −→ 1st bit success rate: ∼50%
2
“Side channel attack: an approach based on machine learning”, 2011,
by L. Lerman, G. Bontempi, and O. Markowitch
Power Trace Alignment with DTW

Motivation
Power traces can be misaligned for several reasons, such as

Synchronization issues between the sampling devices and the
tested hardware
Clock variabilities and instabilities
Intentional countermeasures such as delays and modulations
Misaligned traces ⇒ incorrect/inaccurate correlations ⇒ wrong

classification and useless attacks

Naı̈ve approach: static alignment by time offset
Theoretically:

Theoretically:
a
Use time offset

to align traces
a-

Realistically:

a

a-
Realistically:
a a-
Which offset to use?

Machine-learning approach: adaptive alignment by Dynamic Time Warp (DTW)



Using pairwise alignment in an attack
Training:
1 Acquire power traces
2 Choose reference trace (e.g., arbitrarily or use mean of all traces)
3 Align each trace to the reference trace using the pairwise
alignment
4 Apply training algorithm (e.g., PCA & SVM) to the aligned
traces
Online:
1 Acquire trace from attacked hardware
2 Align trace to the reference trace (from the training) using
pairwise alignment
3 Apply classification algorithm (e.g., PCA & SVM)
Pairwise alignment
Trace x -
i a
Pairwise diff. matrix:
each cell holds difference
6j a qe between two trace entries
H
HH
H x [i] − y [j]
HH

Trace y
H
Pairwise alignment
Trace x -
Alignment path:
get from start to end
of both traces
6
Trace y

Pairwise alignment
Trace x -
1:1 alignment:
trivial - nothing modified
by the alignment
6
Aligned distance:
P 2
= kx − y k2
Trace y

Pairwise alignment
Trace x -
Time offset:
works sometimes, but
not always optimal
6
Aligned distance:
P 2
=?
Trace y

Pairwise alignment
Trace x -
Extreme offset:
complete misalignment -
worst alignment
6 alternative
Aligned distance:
P 2
= kx k2 + ky k2
Trace y

Pairwise alignment
Trace x -
Optimal alignment:
Optimize alignment by
minimizing aligned
6 distance
Aligned distance:
P 2
= min
Trace y

Finding optimal pairwise alignment
Dynamic Programming
A method for solving complex problems by breaking them down
into simpler subproblems.
Applicable to problems exhibiting the properties of overlapping
subproblems and optimal substructure.
Better performances than naive methods that do not utilize the
subproblem overlap.
Dynamic Time Warp (DTW)
Basic DTW Algorithm:

For each trace-time i and for each trace-time j:
Set cost ← (x [i] − y [j])2
Set the optimal distance at stage [i, j] to:
DTW[i,j−1] - DTW[i,j]
 
DTW[i,j−1] 

 
DTW[i,j] ← cost + min DTW[i−1,j−1] 6
 

DTW[i−1,j]  DTW[i−1,j−1] DTW[i−1,j]
Optimal distance: DTW[m,n] (where m & n are lengths of traces).
Optimal alignment: backtracking the path leading to DTW[m,n] via

min-cost choices of the algorithm

Power Analysis with DTW-based Alignment
Use coarse-grained matrices to avoid bad/unreasonable portions:
Drill down by fine graining to approximate the optimal

alignment with quasi-linear time & space requirements
3
“Improving Differential Power Analysis by Elastic Alignment”, 2011,
by J.G.J. van Woudenberg, M.F. Witteman, and B. Bakker
Experimental results:
Compare correlation DPA using 3 alignment methods:
Static: Simple static alignment by time offset
SW: Replace trace entries with avg. of sliding window
Not strictly an alignment method, but simple &
sometimes effective
DTW: Elastic alignment with DTW
3
Static:
SW:
DTW:
DES with stable clock

3
SW:
DTW:
DES with unstable clock

3
Analyzing Non-Cryptographic Leaks
with
Hidden Markov Model

Information & Activity Leaks
Consider secret sequence of activities and leaked information with the

following properties:
Contains information about the secret sequence
Contains noise
Insufficient for directly recovering the secret information
If activities follow known statistical patterns, then an attacker can

“guess” secret sequence from noisy leaks.
Attack: find best hypothesis such that:

1 It matches the leaked data
2 Has high probability according to statistical distribution of
activity sequences

Information & Activity Leaks
Can it work?
Leaked information can be used for more than cryptographic

purposes:
Users are predictable - most activities are similar & repetitive
Internet - common websites and surfing routines
Emails/documents - linguistic models
Passwords - most common password is “password”
Others examples: “querty”, “letmein”, “trustno1”, “dragon”,
“monkey”, “ninja”, and “jesus”.
News services often publish lists of most common passwords of
the year/month
Guess activities/information by detecting “reasonable” usage

patterns from leaked data
A statistical model of user activity profile can be used for this task.
Markov Chain
Stochastic Process: h
x
=?

1
+
- q f - q f - qf
qi
i−2 i−1 i
qi+1 =
?
- h
y
@
q i+
1
h
?
=
@
@R
@ z
Transition probabilities:
Pr [qi+1 =? | qi , qi−1 , . . . , q2 , q1 ]

Markov Chain
Markov Process: h
x
=?

1
+
f
qi
- qi−2m - qi−1m - qi qi+1 =
?
- h
y
@
q i+
1
h
?
=
@
@R
@ z
Transition probabilities (no history):

Pr [qi+1 = ? | qi ] = Pr [qi+1 = ? | qi , qi−1 , . . . , q2 , q1 ]

Markov Chain
Keyboard structure & text auto-complete

Markov Chain

Markov Chain

Markov Chain
- tm - em - xm - tm -

Hidden Markov Model
- hm
1 - hm
2 - hm
3 - hm
4 -

Hidden Markov Model
- hm
1 - hm
2 - hm
3 - hm
4 -
o1f
? f
?
o2 o3f
? f
?
o4

Hidden Markov Model
- hm
1 - hm
2 - hm
3 - hm
4 -
o1f
? f
?
o2 o3f
?
o4 f
?
Transition probabilities: Leak probabilities:

Pr [hi+1 =? | hi ] Pr [oi =? | hi ]

Hidden Markov Model
- hm
1 - hm
2 - hm
3 - hm
4 -
o1f
? f
?
o2 o3f
? f
?
o4
Viterbi Algorithm
A dynamic programming algorithm for finding the most likely
sequence of hidden states, especially in the context of Hidden Markov
models.
Acoustic Analysis of Keyboards
based on paper4
4
“Keyboard Acoustic Emanations Revisited”,2005, by L. Zhuang, F.
Zhou, and J.D. Tygar
based on paper4
4
based on paper4
Typed text:
4
based on paper4
HMM only:
4
based on paper4
HMM & spelling corrections:
4
based on paper4
Password retrieval
4
Acoustic Analysis of Printers
(Pictures taken from

(Picture taken from URL:flylib.com/books/en/2.374.1.27/1/) URL:mindmachine.co.uk/book/print 06 dotmatrix overview01.html)
5
“Acoustic Side-Channel Attacks on Printers”,2010, by M. Backes, M.
Dürmuth, S. Gerling, M. Pinkal, C. Sporleder
sound intensity
# of needles hitting the paper

5
Training:
Feature extraction (split into words, noise reduction, etc.)
Construct DB with (word, sound) pairs
Online:
Feature extraction (same as in training)
For each word:
Sort DB by similarity/difference from recorded sound
Reorder DB by n-gram/word distribution using HMM
Guess printed word as the top candidate from reordered DB
5
5
Further Reading I
Side-channel attacks using machine learning tools:

“Further hidden Markov model “Machine learning in side-channel
cryptanalysis” (2005) by P.J. Green, R. analysis: a first study” (2011) by G.
Noad, N.P. Smart Hospodar, B. Gierlichs, E. De Mulder, I.
“Analyzing side channel leakage of Verbauwhede, J. Vandewalle
masked implementations with stochastic “Side channel attack: an approach based
methods” (2007) by K. Lemke-Rust & C. on machine learning” (2011) by L.
Paar Lerman, G. Bontempi, O. Markowitch
“Side channel attacks on cryptographic “Side channel cryptanalysis using
devices as a classification problem” machine learning” (2012) by H. He, J.
(2007) by P. Karsmakers, B. Gierlichs, K. Jaffe, & L. Zou
Pelckmans, K. De Cock, J. Suykens, B. “PCA, eigenvector localization and
Preneel, B. De Moor clustering for side-channel attacks on
“Theoretical and practical aspects of cryptographic hardware devices” (2012)
mutual information based side channel by D. Mavroeidis, L. Batina, T. van
analysis” (2009) by E. Prouff & M. Laarhoven, E. Marchiori
Rivain “Efficient Template Attacks Based on
“Cache-timing template attacks” (2009) Probabilistic Multi-class Support Vector
by B.B. Brumley & R.M. Hakala Machines” (2013) by T. Bartkewitz & K.
Lemke-Rust

Further Reading II
Trace alignment:
“Recovering secret keys from weak side “Improving differential power analysis by
channel traces of differing lengths” (2008) by elastic alignment” (2011) by J.G.J. van
C.D. Walter Woudenberg, M.F. Witteman, B. Bakker
“Side Channel Analysis enhancement: a “A general approach to power trace alignment
proposition for measurements for the assessment of side-channel resistance of
resynchronisation” (2011) N. Debande, Y. hardened cryptosystems” (2012) by Q. Tian &
Souissi, M. Nassar, S. Guilley, T.H. Le, J.L. S.A. Huss
DangerBakker
Information retrieval from leaked data:

“Keyboard acoustic emanations revisited” “Automated black-box detection of
(2005) by L. Zhuang, F. Zhou, J.D. Tygar side-channel vulnerabilities in web applications”
“Acoustic side-channel attacks on printers” (2011) by P. Chapman & D. Evans
(2010) by M. Backes, M. Dürmuth, S. Gerling, “Current events: identifying webpages by
M. Pinkal, C. Sporleder tapping the electrical outlet” (2012) by S. S.
“Building a side channel based disassembler” Clark, B. A. Ransford, J. M. Sorber, W. Xu, E.
(2010) by T. Eisenbarth, C. Paar, Björn G. Learned-Miller, K. Fu
Weghenkel “Engineering statistical behaviors for attacking
and defending covert channels” (2013) by V.
Crespi, G. Cybenko, A. Giani

Conclusion
Machine learning
Retrieve meaningful information from vast amounts of leaked data.
Machine learning tools/concepts:

Training/testing scheme
Dimensionality reduction with PCA
Clustering/classification with SVM
Alignment with DTW
Predicting/guessing usage patterns with HMM
Side channel applications

Template based power analysis & power trace alignment
Acoustic analysis of keyboards & printers
Conclusion
Machine learning
Retrieve meaningful information from vast amounts of leaked data.
Machine learning tools/concepts:

Training/testing scheme
Dimensionality reduction with PCA
Clustering/classification with SVM
Alignment with DTW
Predicting/guessing usage patterns with HMM
Side channel applications

Template based power analysis & power trace alignment
Acoustic analysis of keyboards & printers
Big Data

Big Data

Big Data

Big Data

Big Data
Not just from security leaks...
Big Data is produced and collected everywhere:

Web data
E-commerce (Amazon, Ebay, etc..)
Purchases at department/grocery stores
Bank/Credit Card transactions
Netflix/Blockbuster/VOD
Social networks (Facebook, Google, etc.)

Big Data
What can be done with it?
Processing Big Data involves several aspect such as:

Aggregation and Statistics
Data warehouse and OLAP
Indexing, Searching, and Querying
Keyword based search
Pattern matching (XML/RDF)
Knowledge discovery
Data Mining
Statistical Modeling

Lecture6 0

Uploaded by

Copyright:

Available Formats

Lecture6 0

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture6 0

Uploaded by

Copyright:

Available Formats

Information Security – Theory vs.

Tel Aviv University

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 1 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 2 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 3 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 4 / 61

Machine learning encompasses tools that perform smart analysis of

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 5 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 6 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 7 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 8 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 9 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 9 / 61

Correlation indicates the correct time and key

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 9 / 61

Correlation indicates the correct time and key

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 9 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 10 / 61

:::: trace ::::

Essentially, power traces are high-dimensional vectors with

:::: trace x ::::

Traces can be analyzed as vectors in the Euclidean space m with `2 R

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 12 / 61

Template assumption: the positions of traces in Rm (with `2 norm)

Correlation is either directly seen or via some internal value

Model power consumption of encrypting plaintext p with key k as a

Attack single trace (with known plaintext p):

Unfortunately, directly analyzing high-dimensional vectors is usually

Common solution - use dimensionality reduction methods and analyze their

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 16 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 17 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 17 / 61

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 17 / 61

tracei [t1 ] · tracei [t2 ]

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 18 / 61

Many datasets (incl. power traces) have a decaying cov. spectrum

Approximate cov. matrix by truncating small eigenvalues from SVD

Points on the line are defined by multiplying ψ ~ by scalars

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 21 / 61

Points on the line are definedkck

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 21 / 61

Points on the line are defined by multiplying ψ ~ by scalars

Covariance matrix has a single eigenvalue k~c k2 and a single

Guy Wolf (TAU) Machine Learning & Side-Channels 2013 21 / 61

6j a qe between two trace entries