Download HumanAut (or SecHCI: Secure Human

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Password strength wikipedia , lookup

Cryptanalysis wikipedia , lookup

Uncertainty principle wikipedia , lookup

Computer security wikipedia , lookup

Subscriber identity module wikipedia , lookup

Secure multi-party computation wikipedia , lookup

Transcript
HumanAut
(or SecHCI: Secure HumanComputer Identification System
against Peeping Attacks)
Shujun LI
Xi’an Jiaotong Univ.
Oct. 2002
A Brief Introduction


Exchange opinions on the definition and
meaning of HumanAut/SecHCI.
Explain the meaning of so-called
“peeping attacks”.
What is HumanAut or SecHCI?


In Prof. M. Blum’s words : HumanAut is such a
system, by which a “naked” human inside a “glass”
house can authenticate securely to a non-trusted
terminal.
In My Opinion
• In real world: against peeping attacks (also called
observer attacks or shoulder-surfing attacks)
• In theoretical world: providing security in
identification (authentication) systems with
untrustworthy devices.
• Another meaning of SecHCI is Secure HumanComputer Interface against peeping attack.
HumanAut/SecHCI: In Graphical
Words
What adversaries
can not see
Computer
Human
What adversaries
can see
Interface
Powerful
Adversary
I am not
secure

me too 
Peeping Attacks: Classification




Passive (Weak) Peeping Attacks: adversaries can
only passively monitor legal users’ responses
Active (Strong) Peeping Attacks: adversaries
control the communication channels and can
disguise themselves as fake verifiers
Hidden Peeping Attacks: adversaries are hard to
be detected (such as hidden cameras)
Open Peeping Attacks: adversaries can be easily
detected by users (such as your friends standing
besides you)
Why Normal Identification Systems
are not OK against Peeping Attacks?


Three Types of Identifications
•
•
•
Knowledge-based – what do you know
Token-based – what do you have
Biometrics-based – what are you
Most identification systems are
•
•
Absolutely insecure to peeping attacks, such as fixed
passwords.
Secure to peeping attacks but require trustable
devices, such as RSA SecurID® card.
Some Solutions of
HumanAut/SecHCI?



Matsumoto-Imai Protocol in EuroCrypt’91:
cryptanalyzed by C.-H. Wang et al. in
EuroCrypt’95
Matsumoto Protocols in ACM CCS’96: can
only resist O(v) observations, where v is the
size of each challenge question.
Hopper-Blum Protocols in AsiaCrypt’2001:
the best ones from the viewpoint of security
till now, but better usability is wanted.
Two More Points on
HumanAut/SecHCI

CAPTCHA is useful to relax the security requirement on
online attacks, since humans can only carry out attacks
with much lower speed than computers.
•

So, can we use the same challenges to realize identification
and CAPTCHA simultaneously?
HumanAut/SecHCI can be extended as tools of AVT –
age-verification technology.
•
•
AVT is used to protect kids from improper (especially
pornographic) materials on computer/Internet.
The responses to challenges should be designed to be
almost impossible (i.e., very very difficult) for kids (even
teaching them how to use the protocols is difficult), but
capable for most adults (usability can be relaxed, for
example, it is still OK if some training is needed).
Our Ideas on HumanAut/SecHCI



Introduce our basic ideas on the design of
HumanAut/SecHCI.
Hope that Prof. Blum can point out
problems in our proposals, and give some
suggestions on our future research.
Since ALADDIN Center is doing the best
research on HumanAut, Harry and I would
like to make joint research with Prof.
Blum’s group.
How Does a Peeping Attack
Works: I


Conceptually, let us write a
challenge-response pair as an
equation f(c(P), P)=r, where P
is the password with k secret
parameters.
Assume an attacker A has
observed n challengeresponse pairs, then he gets
an equation system with k
unknown variables, which is
shown in the right side.
 f (c1 ( P ), P )  r1



 f (ci ( P ), P )  ri



 f (cn ( P ), P )  rn
How Does a Peeping Attack
Works: II

Attacks: apparently, when n is large enough, it is
possible for A to exactly or numerically solve this
equation system to get the password P.
•

In Matsumoto Protocols [ACM CCS’96] and Hopper-Blum
Protocols [AsiaCrypt’2001], the equation system is linear
and k independent equations are enough to unique solution.
Uncertainty: frustrate the above attack
•
•
•
If ci and/or ri are uncertain, the solution becomes
probabilistic.
Uncertainty can be exerted on the left side (challenge) or
right side (response).
In Hopper-Blum Protocols, uncertainty is provided on the
right side by introducing intentional errors.
Some Design Factors?

Uncertainty is the basic tool to frustrate peeping
attacks.
•

Balance is important to provide “effective” uncertainty in
HumanAut/SecHCI, otherwise the difference can be useful
for attackers to clarify uncertainty
•

Intentional response errors and/or redundancies may be useful
to enhance security. A problem of uncertainty is that usability must
be sacrificed to some extent.
In Hopper-Blum Protocol 1, the insecurity to active peeping attacks
is partially caused by the fact of ≠1-.
Visual/graphical implementations may be helpful to
enhance usability, and the security against dictionary attacks
(it is much difficult to compose a graphical dictionary than a
textual one).
New Ways to Uncertainty: I

A generalized version of Hopper-Blum Protocol 1 with balance
property (but can be naturally extended to any protocol)
• C=>H: c1, c2
• H=>C: r1, r2, where only one response is right and another is intentionally wrong with
private (and balanced) coin-toss
Repeat the above steps for m rounds


•
Security Analysis
•
•
•
The coin-toss should be really private and balanced.
The success probability of guessing n right responses is 2-n/2, which
should be small enough to provide acceptable security.
More attacks?
Usability Analysis
•
•
The extra wrong responses (half of all ones) make the usability worse
than protocols in which all responses are right.
It is generally hard for humans to make really good coin-toss.
New Ways to Uncertainty: II

A general model to introduce uncertainty
Computer
Human
Password
Time-Variant
(Pseudorandom)
Source
Challenge
Unique
Response r’
Generic Challenge-Response Protocol
Also, can be a
Probabilistic (or
Fuzzy) Map
Balanced
Multiple-to-One
Map f ()
Mapped Response r
One round of the proposed model for SecHCI protocols
Some Points on the Model



The multiple-to-one (or probabilistic, fuzzy)
map should be human-executable (adultsexecutable for AVT).
Such a map can be also considered a
classifier, which outputs the class number
for each input. Here, each class should
contain at least two elements to make the
classifier be a multi-to-one map.
Apparently, r should have at least 2 different
values, so r’ should have at least 4 values.
An Example Protocol



Given a set O containing n objects, Password
P is a k-size subset of O, where k>4.
Given a subset C of O, similarity of C is
defined as the size of PC and denoted by
Sim(C). Here, C is also called a challenge-cell.
One round of the protocol is as follows:
•
•
Computer=>Human (Challenge): C1,C2,C3,C4
Human=>Computer (Response):
 Sim (C1 )  Sim (C2 )   (Sim (C3 )  Sim (C4 )  mod 4 
r
 0,1

2


How to Realize Balance in the
Proposed Protocol: I


To make r balanced, Sim(C1)+Sim(C2) and
Sim(C3)+Sim(C4) should be distributed in the
set {0,1,2,3} uniformly.
A simple way to realize the uniform distribution is
to generate C1,C2,C3,C4 in all subsets whose
similarities are 0,1,2,3 with 1:1:1:1 ratio (Rule A).
• But such a way causes attacks based on partiallyknown password.
• When a passive attacker get k’3 elements in P, he can
•
get some challenge-cells whose similarities are 3. These
cells can reveal which elements are not included in P.
As a result, one of C1,C2 (and C3,C4) must be generated
at random in all subsets (Rule B).
How to Realize Balance in the
Proposed Protocol: II

For challenge-cells generated with Rule A, the
occurrence probability in challenges of
elements in P and the probability of the
elements not in P should be balanced, too.
•
•
Since n should be large enough to provide security, we
prefer to using fixed-size challenge-cells to reduce
the number of displayed objects in screen.
Assume the size of each challenge-cell is l, from the
balance of the occurrence probability, 1.5n=kl, where
1.5 is the mean similarity of all challenge-cells
generated with Rule A.
A Textual Implementation



Password: P={m, a, n, u, e, l, b, u} (which
means Manuel Blum )
Challenge:
•
•
•
•
C1={c, m, q, z, *, i, k, u} (Sim=2)
C2={a, r, &, i, e, 2, k, l} (Sim=3)
C3={g, r, o, d, f, !, q, w} (Sim=0)
C4= {p, e, b, y, h, j, ., s} (Sim=1)
Response:
 (3  2  0  1) mod
r
2

4
 (3  2  0  1) mod
  1 or  
2
4
  0
A Graphical Implementation
Password
Challenge
Sim=3
Sim=0
Response
 (3  2  0  1) mod
r
2

Sim=2
Sim=1
4
 (3  2  0  1) mod
  1 or  
2
4
  0
Security Analysis: I


For passive peeping attacks, the proposed protocol
seems secure with 2n complexity.
For active peeping attacks, the adversaries can
successfully find the challenge-cells generated with
Rule B, since the response has unbalanced
relationship with the similarities of such cells. But the
multi-to-one map makes the right similarity uncertain,
and the success probability is not greater than p-n,
where
 p1
p3
p3
p2
p4
p2
p1
p4 

.
p  max 
,
,
,
,
,
,
,
 p1  p2 p1  p2 p3  p4 p3  p4 p2  p3 p2  p3 p1  p4 p1  p4 
Security Analysis: II


Attacks?
• Prof. Blum’s criticism
• Our further investigations
Modifications?
• Prof. Blum’s suggestions
• Our further investigations
Usability Analysis: I

Because of 1.5n=kl, k and l will be a little large since
n must be large enough.
•

Generally, for active peeping attacks, assume p=0.75 (an
approximate value), n150 for O(260) attack complexity,
n200 for O(280) complexity, and n250 for O(2100) complexity.
Too many symbols must be displayed on the screen:
4l for each challenge.
•
•
Text/icon-based implementation will be useful to relax this
problem, where “icon” means graphics with small size.
Drawing-based implementation may be another candidate
to solve this problem. A typical idea is reported in DAS
graphical passwords [USENIX Security ’99].
•
Assume elements in O are different strokes in a mn grid, it is
possible to display multiple strokes in a same grid, which save
display space dramatically.
Usability Analysis: II

The consuming time for each identification is
t0*m, where t0 is the mean time for one round
and m is the round number.
•
•

The larger k, l, m are, the larger the time will be.
A textual implementation shows that the consuming
time is rather great, so graphical implementations are
needed to solve this problem.
More Problems and Solutions?
• Prof. Blum’s criticism and suggestions
• Our further investigations
More Protocols?

In fact, based the idea of introducing
uncertainty in responses by multiple-to-one
map, many different protocols can be
constructed.
•

Hopper-Blum Protocol defined on {0,1,2,3,4,5,6,7,8,9}
[AsiaCrypt’ 2001] may also be modified.
Extended models?
•
Can we generalize the model to introduce uncertainty
in the challenge side and both sides?