Cryptography Reference
In-Depth Information
The sophistication of the adversaries in our threat model is determined in terms of
their ability to collect reconnaissance on the internal structure of the alert distribution
mechanism and focus their attack.
Thus, at a high level, we distinguish between two
types of adversaries:
- Adversary with random attack : Unsophisticated adversaries who can inhibit com-
munication to/from randomly selected nodes. The fraction of nodes they can bring
down depends on the level of resources available to them.
- Sophisticated adversaries, who exploit knowledge of the system structure to target
nodes such that they maximize the impact of their disruption. We further consider
two sub-types of such adversaries:
1. Adversary with targeted attack: Attackers that know and exploit the high-
level structure of the network topology. Such attackers, for example, know
the identity of and target the super-nodes or other, relatively “fixed” important
nodes in the system.
2. Adversary with degree dependent attack: More powerful adversaries that
somehow have detailed topology information about a large part or all of the
distribution mechanism. Such knowledge includes, for example, the complete
connectivity graph of the participating nodes (or a large fraction thereof).
For all type of the above schemes, selected victim nodes are taken out from the system
as a consequence of the attacks.
4
Implementation
For our evaluation, three different alert distribution systems were implemented on Over-
Sim [3] network simulation framework. Here, we describe how we implemented the
simulation modules. We first talk about the design choices for the signaling channels
and the various reliability parameters that we explored. Then, we cover communication
models considered for alert distribution systems.
4.1
Control Channel
Centralized System. In the case of centralized control, we employed a super node (SN)
network. Among many configuration parameters [27] for the SN network, we carefully
identified the ones that affect the robustness of the overall network: the size of sub-
network ( cluster size ) and number of super-node replicas ( k-redundancy ). The cluster
size was tested using a range of different values. The same holds for k-redundancy .
However, in our graphs, we present only the case where k-redundancy is two. We did
so because other values of k-redundancy do not notably change the system's behav-
ior beyond the one captured by the graphs. We configured the rest of the parameters
unchanged as these parameters have an effect only on the network performance.
Distributed System. For distributed control, we chose Chord [25] to implement a de-
centralized alert notification system. Chord was selected for two reasons. First, Chord's
 
Search WWH ::




Custom Search