Combining Observables

Real-world analytics often incorporate multiple observables, so we now turn our attention to scoring groups of observables, in particular the evaluation of boolean expressions in analytics.

Robustness as a Metric

Robustness is a metric that we have devised to give a relative score for the difficulty an attacker would have in evading a candidate analytic when executing a given technique. Robustness is determined by the lowest level that an analytic contains (according to our levels of robustness) that could be evaded by an adversary and lead to a successful attack. Thus, the highest robustness level (7) requires the adversary to make a costly change to their TTPs, whereas the lowest level (1) requires only quick and inexpensive changes to.

Evaluating Robustness

Let R(X) be the Robustness Level for a given analytic.

  • If the analytic contains a single observable O, then the following rule applies:
    • R(X) → R(O). That is, the robustness of the analytic evaluates to the robustness of the observable.

  • If the analytic contains multiple observables, then the following boolean logic applies:
    • R(A AND B) → MIN(R(A), R(B)) - With the boolean operator “AND”, the adversary only needs to evade either A or B, which makes the robustness equal to the lesser of the two observables.

    • R((A AND B) | A) → R(B) - The robustness level of A and B predicated on observing A is equivalent to the robustness level of B, since observing A is a given in this context.

    • R(A OR B) → MAX(R(A), R(B)) - With the boolean operator “OR”, the adversary needs to evade both A and B, which makes the robustness equal to the greater of the two observables. Note a special case where two observables at level 4 happen to cover all possible implementations, then that would raise the boolean OR expression to level 5.

    • R(NOT A) → R(A) - The robustness level of NOT A would be equivalent to the robustness level of the observable A itself, since the detection focus is still at A’s level of robustness.

Precision and Recall

Although precision and recall are outside the scope of this project, we briefly touch on how these metrics are affected by boolean expressions:

  • OR and IN operators expand the search aperture, which increases recall (but might tradeoff some precision)

  • AND operators shrink the aperture, which decreases recall (but might improve precision)

We do not try to claim that maximizing recall or maximizing precision makes for the best analytic, as it is highly dependent on what the objective of the analytic is and the environment it is being used in.