scyther/src/notes.txt

77 lines
4.9 KiB
Plaintext
Raw Normal View History

2004-04-23 11:58:43 +01:00
- best heuristic currently for -t8: ignore sendsfirst, just look at the
previous event and continue at that run.
- termPrint must currently be encapsulated in a math environment to work in
latex output, I wil change this later to some ensure mathmode construct,
probably with something like latexTermPrint.
- Repository to apache2 webdav.
- The trace incremental search should start at something like 'distance of
first claim + SOMECONST * runs'. SOMECONST relates to the partial-order
induced trace prolonings.
- For now, CLP is non-working and not checked. Reconsider this later.
- Because the properties checked are related to the partial order reductions,
it makes sense to check 'all secrecy claims' or 'all synchronisation claims'
at once.
- Once an attack is generated, we know the actual scenario and the trace
length. We might re-use this information to generate the shortest attack
using no partial order reduction. Note that choosing a wise method of
finding that is a mathematical issue: searching -t1 is exponential, so how
do we step trough these traces in order to find the shortest trace in the
quickest way? This can be determined under some (statistical) assumptions.
- The incremental check can be done much easier, by just retrying the main
loop explore with different parameters. It keeps everything nicely intact.
We must check for fails however, and how pruning is affected.
- Incremental check can at least be done in two ways: either by limiting the
number of runs (big axe method) or by limiting the trace length (nice).
There is a drawback to trace length only: the superfluous sends of the
algorithm in a scenario with many runs might interfere with the usefulness
of short traces. Maybe some heuristic that reduces the runs automatically is
required. (e.g. 'number of runs is at most MaxTraceLength/MaxRoleLength' or
something like that). Do analysis.
- If synch checking is expensive, introduce a limit (for -p0 stuff).
Basically, after the limit (if -p0 is set), we stop checking synch claims.
This then is default behaviour, to be overridden or changed by a switch.
- clp does not yet consider the matching for variables that have not been
instantiated. This *might* lead to some improvement, e.g. when there is only
one candidate for the variable. In general, I don't think it will help.
- match_clp.c 1.25 introduced the variable_in_invkey brancher, which hasn't
been thoroughly tested yet. Please do test.
- Brutus is slower than -t4. CLP is much, much faster however than both.
- It is very important to have variables in the run agent list, because it
dramatically decreases the complexity of the system, as opposed to just
creating runs for all possibilities. However, this explicitly introduces the
clp problem of var encryptions (see above).
- Given originator assumptions and such, we might want to implement the
compiler, that for non-typed matching, it will ignore all type definitions
in the code (as in, putting them in stype, but still considering Function).
Then, for 'any agent' we can use stype silently. The modelchecker should
then always consider the types with -m0 and -m1.
- In a paper concerning the CASRUL compiler, they mention associativity for
the equality relations. Note that CASRUL restricts the intruder actions,
and sometimes uses approximations.
- MS and CE never mention associativity as a restriction.
- For now, clp stuff will use pairing and simple mgu. Maybe a -m3 switch will
add the associative variant later.
- MALLOC_TRACE is the environment variable that triggers mtrace() output.
- Several memory leak detectors are still in place, they can all be located by
searching for "lead" or "diff" and destroying the code.
- knowledgeAddTerm might be improved by scanning through key list only with
things that are newly added.
- Associativity and unification don't play together well; consider other
algorithms. Maybe a simple explore (x,y),z and explore x,(y,z) will do, e.g.
take any split point for both sequences.
- POR (-t3) is actually worse than -t2: sends first. After some pondering I
now know why. It's because simultaneous reads yield traces in which there
are attacks that are also in the ones where we have R0S0S0;R1 (i.e. is thus
equivalent with R0;R1;S0S0)
- Knowledge is monotonous. So, if we say 'R#0' cannot be triggered by some
knowledge, we can just use a pointer for it. If we do want to trigger it, we
need to compare two knowledge pointers. As they don't have a unique
representation, it is more difficult. For now, I will just forbid specific
branches of the read when a message is in some previous knowledge, which is
more precise.
- Comparisons with Casper are in order.
- It is a good idea to have any number of input files, and parse them all, to
combine scenarios. A command should be "require <file>", because that would
enable scenarios to load the required role definitions. It's require instead
of include, because of possible multiple occurrences.