Security Concerns:
Data leakage
Data handling on the provider (Correct storage
procedure, correct deletion)
Correct Computation
Legal issue
…
Attackers:
Outsider
(Ristenpart et Al. - 2009)
Ristenpart et Al. studied data leakage on
Amazon EC2.
Threat model:
There is a single trusted cloud maintainer that
manage a public multi-tenant cloud.
The attacker is a normal.
EC2 Internals
EC2 uses Xen hyp. Domain0, it manages
guest image, physical resources accesses ecc…
In EC2 Dom0 route packets to VMs and
reports itself as hop in traceroute.
Ec2 offers five instances: small, medium,
large, … (small is a single virtual core…)
Network: availability zones do not share
Different VM share the same Ph.Inf. It is
possible for an attacker to achieve co-residence:
Since Dom0 appears in trace-route there is a
“free” and deterministic co-residence check.
There are some bias in VM assignment
The strong placement locality can be
exploited by an attacker to achieve co-residence.
Cross-VM information leakage:
The contention on buffers may be used as:
Covert Channel: [Xu et Al-2011] carefully studies the
achievable bitrate of L2 cache contention using different protocols.
Estimate the load on the target VM-machine
[RTSS09] teach us that blind trust in not
the best option.
We assume that cloud provider/s is/are
not trusted:
Two adversaries:
Honest-but-curious Malicious (byzantine)
Avoid data leakage
The only way to ensure privacy of
outsourced data is to encrypt them but:
The naïve encryption rule out any form of
computation over data.
Over the years have been developed many
techniques to overcome that:
Partial-Homomorphic: RSA (multiplicative), Paillier
(additive).
Specific computation only: Searchable data
encryption, Order preserving ecc…
Homomorphic Encryption
Breakthrough in Cryptography:
Gentry STOC-09 shows how to achieve
fully-homomorphic encryption using ideal lattices-bootstrap theorem.
In June 2010 D.G.H.V. shows how to
achieve Fully-H.E. over integers
Fully H.E. over the Integers
We have a circuit C \in Ce and a function
evaluate, a scheme (Dec,Enc,Evaluate) is homomorphic (w.r.t Ce) if given a tuple of Ciphertexts c=(c1,..cn) we have:
Dec[sk,Evaluate(pk,C,c)]=C(m1,…mn)
To rule out trivial scheme there is the
compactness property:
There exist a fixed polynomial bound b(n) so
that for any condition (sk,pk, C, c) the size of Eval[pk,C,c] < b(n).
Steps to achieve a non trivial Fully-H.E:
Find a somewhat homomorphic private
encryption scheme that respect some conditions.
Turn the scheme in circular secure public key
scheme
Use the bootstrap theorem [Gentry-09] on the
Let us start with a simple private key
encryption scheme:
KeyGen:
Encrypt(p,m):
Problems
For each call to Evaluate we have:
Ciphertext Grows: double the bit each
multiplication violate compactness
Noise Grows: for each addition and
multiplication the terms that are not multiple of p grows. Noise > p/2 violate correctness.
Public Key
Approximated-GCD problem
Given an oracle for a randomly
chosen p output p.
Given an adversary A that breaks the
presented scheme in p.t. with advantage e it is possible to build an adversary A’ that breaks A-GCD in p.t. with probability p(e).
Win == get p
A Q LSB Oracle
z1=zq1p+zr1 c=(m+zb+S*Pk) a Binary GCD (z=qp+r,q) Pk:{x0,x1,…x,_n} z2=zq2p+zr2 zb
m <- {0,1}
S<-{0,1}^{n}
Bootstrap - intuition
If the scheme is able to evaluate is own
decryption procedure C_d then it is possible to use C_d to “decript” a
E(m,Pk1) while it is encrypted under key Pk1 using the E(k1,Pk1).
Performance is (but for how long?
No-bootstrap Result-2012) the main drawback of Fully-H.E.
(To achieve circuit privacy we need garbled circuits)
Other problem can be solved in a more
efficient way (or only using) other techniques
Yao Garbled Circuit (1986)
Good introduction: http://www.cs.illinois.edu/class/
fa09/cs598man/slides/AC-F09-Lect16-yao.pdf
Using Garbled Circuit it is possible to
achieve circuit privacy and secure two party computation.
Fairplay
Data Handling
Many services offer the capability to store
data on cloud (Amazon S3, SkyDrive,
Dropbox…), how we can ensure that the cloud maintainer handle this data
properly?
Two issues:
Integrity
We have a huge dataset (>10 TB) and we
want to outsource it. Since we do not trust the maintainer we want to devise an
integrity mechanism.
Naïve: compute MAC on dataset,
drawbacks?
Solution:
Remote Data Checking using Provable Data
Server
Client
< F={b1,b2,…..,bf}, T > Client sends F and T to S
Server Client Verification Phase Init Phase Request: O(1) Response: O(1)
MR-PDP
It is possible to extend PDP to k-replicas :
Naïve way 1: use the same PDP k times.
Vulnerable to coalition
Naïve way 2: use different PDP using k
different enconding.
Expensive O(nk*Tagtime)
Idea create k different file that are
“related” and securely obfuscated.
For each replica we pick a random u.
Using a PRF F for each original block b_i we create b’_i=b_i+F(u|i)
Data deletion
Law impose to securely delete sensitive
information. (Medical Records- Credit Card Number)
A way to do that is to securely wipe
(overwrite) data.
This is not viable on cloud storage:
This issue has been addressed recently:
FADE (Tang et Al. 2010)
FadeVersion (Rahumed et Al 2011). ADEC
(Tezuka et Al. March 2012)
The rationale behind all this system is
ADEC
E(F1,k1) E(F2,k2) E(F3,k3) E(F4,k4) E(F4,k4) E(F5,k5) E(F6,k6) V2 V1 V1 metadata E(<k1,k2,k3,k4>,kv1) h1,h2,h3,h4 V2 metadata E(<k4,k5,k6>,kv2) h1,h2,h3,h4 Cloud: S3, SkyDrive, iCloud, ...Oblivious m-Assignment
Assignment Algorithm are fundamental in
many field: Resource Sharing, Channel Assignment.
Cloud maintainers know the assignment of
resource to clients:
Is it possible to coordinate concurrent entities
such that each one knows is resource but do not know the other assignments?
fairness?
Model
Problem Definition
Oblivious assignment with m Slots (O-mA)
is specified by the following properties:
Unique Assignment (Safety)
Lockout Avoidance (Liveness)
Oblivious Assignment (Obliviousness): if a slot
r_j is assigned to an honest process p_j no
other process is deterministically aware of this assignment
Strong O-mA:
Solvability Issues
Permission algorithms are not suitable for
solve O-mA
In permission algorithms a process ask if
it is safe to access CS…
Perpetual Circulating Token: The trivial
Ensure Fairness: Rotating Leader can
enter in CS.
The other processes must have a non zero
probability to gain CS.
Must be not possible to distinguish (in
Assignment Phase
1
2
E(t1,PPk) E(t2,PPk)
...
1 2 3 4 (pm,PPk) E(tx,PPk) ... E(t_{n-1},PPk-2) n E(t1,PPk-2) ... E(t_{x-1},PPk-2) E(t_{x+1},PPk-2)
1
2
(p_2,PK2)
E(tx,PPk)
... (p_3,Pk3)
1 2 3 4 E(tx,PPk) ... n (p_3,Pk3) (p_{n-1},Pk_{n-1}) (p_2,Pk2) tx
Rel. Phase
1 2 ... E(0,PPk-2) E(0,PPk-2) tx E(b,PPk-2) ... E(0,PPk-2) E(b,PPk-2) b= released? xor1 2 3 4 ... n 0 0 b ...
What is the number of winner ticket
20 40 60 80 100 b 10
20 30 40
w=40 w=30 w=20 w=10
0.4 0.6 0.8 1.0
w=40 w=30 w=20 w=10
20 40 60 80 100 r 0.2 0.4 0.6 0.8 1.0 p=0.95 p=0.5 w=50 w=40 w=30 w=20 w=10