Check out the new USENIX Web site. next up previous
Next: Role Classification Up: Model Previous: Model


Defining Similarity

We use connection behavior as a basis for host grouping, because that information is easily available by just monitoring the network. To group hosts, we need to define similarity in a way that captures the extent to which pairs of hosts establish connections to the same set of other hosts. We start by defining similarity between hosts as a function of the number of common hosts with which they communicate. Intuitively, hosts that share the same logical role communicate with similar sets of hosts. A connection is a pair consisting of a source host address and a destination host address. The connection set of a host, C(h), is the set, $\{ a \mid a \in I$ and there is a connection between h and a}. If $h_1 \in C(h_2)$, then $h_2 \in C(h_1)$. We define the relation neighbor(h1, h2) to be true if and only if h1 = h2 or $h_1 \in C(h_2)$. For later use, we extend the definition of neighbor to groups by defining neighbor(G1, G2) to be true if and only if there exists a host $h_1 \in G_1$ that is a neighbor of another host $h_2 \in G_2$. We can use the notion of a connection set to provide a simple definition of similarity:
\begin{displaymath}
\mbox{\emph{similarity}}(h_1, h_2) = \vert C(h_1) \cap C(h_2)\vert
\end{displaymath} (1)

That is to say that similarity(h1, h2) is equal to the number of neighbors that h1 and h2 have in common. We are now in a position to specify the requirements of a grouping algorithm. Given a set of hosts, I and a similarity threshold, Smin, it must find a partitioning, P, of I that is maximal with respect to avg_similarity and Smin, i.e.,
  1. P respects avg_similarity. This constraint guarantees that each host is within the group with which it has the strongest average similarity.
  2. For all $h \in G$ and $G \in P$, avg_similarity(h, G) $\geq S_{min}$. This requirement guarantees that each host in a group is sufficiently closely related to every other host in the group, thus ensuring that groups are not too large.
  3. There is no other partitioning P of I that meets the first two requirements and has a larger average group size. This guarantees that groups are not too small.
This specification is independent of the definition of avg_similarity. For some networks, such as the one represented in Figure 1, the above definition of avg_similarity yields excellent results. However, for others a slightly more complex definition works better. We present such a definition in Section 4.2.


next up previous
Next: Role Classification Up: Model Previous: Model
Godfrey Tan 2003-04-01