![](https://csdnimg.cn/release/download_crawler_static/16502251/bg5.jpg)
China Communications • April 2019
155
sharing resources (cache-based attacks by
monitoring different cache positions).
3.3 Modeling information leakage
After describing the side-channel attacker and
the client in cloud, we set up a unified infor-
mation leakage model from the client to the
adversary, which can cover a wide variety of
co-residency attacks in practical clouds. Apart
from three key dimensions in [20], we identify
an additional key dimension that will influence
information leakage in side-channel attacks.
We explain these in figure 1 above where we
have two clients: Blue (A) and Red (V) with
3 and 2 VMs respectively, while each Blue’s
VM has different numbers or types of secrets.
Since we have no prior knowledge about the
identification of the adversary, we assume that
any client could be an adversary. In this
case, we assume the Red as the adversary and
the Blue as the client, although vice versa is
also possible. This figure shows the placement
of their VMs across three epochs across 4 ma-
chines with 2 VM slots per machine.
cumulate information from the side channel
across time (epochs, denoted as ∆T) when
co-resident with any target VM. For exam-
ple, if the adversary is co-resident with its
target at time T
1
and T
3
, but not T
2
, the in-
formation gathered during T
1
and T
3
can be
combined.
Ø Information replication across different
VMs of the target: For different VMs host-
ing the same secret belonging to the same
target, co-residency with these VMs will
lead to sum of leakage bits derived from
each of these VMs. While this might lead
to redundant bits of that secret, we follow
the principle as considering an adversary as
powerful as possible.
Ø No collusion across adversaries: It is as-
sumed that creating a new client identifier
(e.g., a verified credit card) will cost so
much that Sybil attacks for collusion [30]
are not possible.
Ø Information collation across different
VMs of the adversary: Considering that
the adversary might launch multiple VMs
for attack, we assume a stronger adversary
model where the attacker is capable of
combining information of the same target
derived by its different VMs. With this as-
sumption and the previous one, we focus
on a very powerful threat model where we
conservatively assume that the adversary
gets unique bits per epoch and per VM,
although in reality, an adversary may get
duplicate/redundant information across
epochs or among different attacking VMs.
Under this practical assumption, we unify
the cases of C and NC in the threat model
of [20].
Ø Paralleled attacks against different se-
crets in co-resident target VM: Since we
assume that a single client VM can host
different secrets, an adversary VM co-locat-
ed with it can attack all secrets in parallel.
This model is more powerful but practical
since different secrets can be leaked by dif-
ferent types of attacks (memory-deduplica-
tion-based [22] or DRAM-based [31] side
channels) or by utilizing different parts of
Fig. 1. Co-residency between different tenants.
Table I. Leakage of Different Secrets to the Possible Adversary.
A-VM-1 A-VM-2 A
s1
vul
3
1
V
vul
2
1
V
vul
5
1
V
s2
vul
1
2
V
vul
1
2
V
vul
2
2
V
s3
vul
2
3
V
vul
0
3
V
vul
2
3
V
Machine
M-1 A-VM-1
M-2
M-3
M-4
V-VM-1 A-VM-1 V-VM-2 A-VM-1 V-VM-2
Epoch-1 Epoch-2 Epoch-3
V-VM-2 V-VM-1 V-VM-3
A-VM-2 V-VM-3 A-VM-2 A-VM-2 V-VM-1
V-VM-3
Secret s s
V VM−−1
= { 1, 2}
,
Secret s s
V VM−−2
= { 1, 3}Secret s
V VM−−3
= { 1}
.
万方数据