Skip to content

A quantitative information flow analysis on PDGs based on a data flow analysis

License

Notifications You must be signed in to change notification settings

parttimenerd/nildumu

Repository files navigation

logo

An experimental quantitative information flow analysis for a while-language with functions, arrays and a C like syntax. Example programs can be found in the examples and the eval-specimen (file ending .nd, from the evaluation) directory.

TL;DR: To reproduce the results of the paper use the docker image parttimenerd/nildumu:

# run the small version of the evaluation, prints the evaluation result table
# takes 12 minutes on an Intel i5-8500 using 9GiB RAM and two cores
# `--programs small` excludes the longer running E-Voting benchmarks
docker run -i parttimenerd/nildumu ./evaluation --programs small --runs 1
# analyze a program (with inlining level 32) from standard in
docker run -i parttimenerd/nildumu ./run << EOF
    input int h; output int o := h | 1
EOF
# analyze a program with a different INLINING level
docker run -i parttimenerd/nildumu ./run --handler 'handler=inlining;maxrec=INLINING;bot=summary' << EOF
    input int h; output int o := h | 1
EOF

Tests

Usage

Nildumu can be used either via its command line application ./run, via an editor ./gui or via the evaluation script ./evaluation (be sure to run ./update_eval before).

Command Line Application

A command line that can be used to run programs (passed either via standard in or via a passed filename).

Usage: ./run [-hV] [-tp] [--algo=<algo>] [--handler=<handler>] <programPath>
Run nildumu on the command line. Example programs are given in the folders
'eval-specimen' and 'examples'. The syntax of the paper is accepted too.
      <programPath>          program file to analyze, or '-' to read from
                               standard in. Nildumu files usually end with '.
                               nd'.
                               Default: -
      --algo=<algo>          Used leakage computation algorithm, default is
                               OpenWBO PMSAT based
                               Default: Open-WBO GL PMSAT
  -h, --help                 Show this help message and exit.
      --handler=<handler>    Method invocation handler configuration, for
                               paper: 'handler=inlining;maxrec=INLINING;
                               bot=summary'
                               Default: handler=inlining;maxrec=32;bot=summary
      -tp, --transformPlus   Transform plus into bit wise operators in the
                               preprocessing step
  -V, --version              Print version information and exit.

You can also run this command inside the docker image (build it via docker build -t nildumu ., run it via docker run -it nildumu) or use it directly, for example:

# directly (or in docker image)    # does only require a JDK >= 8, maven, curl and unzip
./run examples/laundering_attack.nd
# via docker image from dockerhub
docker run -i parttimenerd/nildumu ./run examples/laundering_attack.nd

For convenience, vim, emacs and nano are installed by default.

Evaluation

Use the docker image from docker hub, as stated before: The evaluation in the paper used an Intel Xeon Gold 6230 CPU with 40 cores (80 logical cores) and 512 GiB of RAM. But the benchmarks can also be run on machines with a CPU like the Intel i5-8500 (6 cores, only two are actually used) using 9GiB RAM for the evaluation and the docker container:

# run one single iteration, this takes 12 minutes
# and excludes the longer running E-Voting examples
docker run -i parttimenerd/nildumu ./evaluation --runs 1 --programs small

# to run all benchmarks, pass `--programs all` (takes around 3 days)
docker run -i parttimenerd/nildumu ./evaluation --runs 1 --programs all

# to obtain the "exact" leakages for the E-Voting and Smart Grid programs, run
docker run -i parttimenerd/nildumu ./evaluation --runs 1 --programs all --tools exact

The evaluation commands print the evaluation results.

Running the evaluation without docker is also possible, refer to the Dockerfile for required packages. The evaluation of nildumu alone runs without installing additional software, only the usage of dsharpy (which includes an implementation of ApproxFlow and relationscutter), which we compare with nildumu, requires additional software (the version of the SEFM21 paper has is the commit 2fcf12994d).

Nildumu Editor UI

editor An editor front-end for nildumu which gives additional information. Run it via ./gui, requires Java >= 8

Language

The language is loosely C-based. See the beforementioned examples and the documented grammar at src/main/antlr4/Lang.g4.

UI

The following describes the UI (launch it via ./gui). All configurations and inputs are stored continuously in the gui.config file, they are brought back after a restart of a program

  • the top combobox allows accessing different examples from the examples folder and to store the current program therein, using the "Save" button
  • the text input allows to input the program, it supports basic auto completion, syntax highlighting, code folding, and syntax error highlighting
  • the checkbox labeled "Auto Run"
    • if checked, the input program is analyzed continuously
    • only use this for small programs with short analysis times
  • the checkbox labeled + → &!|^
    • if checked, all + and - operators will be replaced by their respective bit operation networks
    • this yields to a far worse performance
    • there is no difference in the leakage computation
  • the combobox labeled Output
    • MIN: only update the Leakage table
    • WONR: also update the Variable values table
    • ALL: also update the Node values table and output debug information on the console
  • the button labeled "Run"
    • run the analysis, changes to during the run of the analysis
  • the button labeled "Stop"
    • abort the current run of the analysis
  • the big combobox below
    • it allows the selection and configuration of the method handler that handles the method invocations
    • some example configurations are already included
    • it uses as syntax the standard Java property syntax, using ; instead of line breaks
    • handlers are selected using the property handler
      • if this misses and the configuration string just consists of a single identifier, then this identifier is considered as the chosen handler
    • the basic handler (all handler in the paper)
      • this is the default handler, that just connects returns a return value for every function call in which every bit depends on every parameter bit
      • this yields to maximal over-approximation, but it is fast and therefore used as default handler to gather some kind of basis for other handlers
    • the inlining handler
      • A call string based handler that just inlines a function.
      • If a function was inlined in the current call path more than a defined number of times (maxRec), then another handler is used to compute a conservative approximation
      • properties
        • maxrec: default is 2
        • bot: the handler for the overapproximation, just a handler configuration, as the current one
          • allows to chain handlers, it might be useful to use the summary handler here
          • default is basic
    • the summary handler
      • A summary-edge-based handler.
      • It creates for each function beforehand summary edges:
        • these edges connect the parameter bits and the return bits
      • The analysis assumes that all parameter bits might have a statically unknown value.
      • The summary-edge analysis builds the summary edges using a fixed point iteration over the call graph.
      • Each analysis of a method runs the normal analysis of the method body and uses the prior summary edges if a method is called in the body.
      • The resulting bit graph is then reduced.
      • It supports coinduction (mode=coind) and induction (mode=ind, the default)
      • Induction starts with no edges between parameter bits and return bits and iterates till no new connection between a return bit and a parameter bit is added.
        • It only works for programs without recursion.
      • Coinduction starts with an over approximation produced by another handler (bot property) and iterates at most a configurable number of times (maxiter property), by default this number is 2147483647 (the maximum number of signed 32 bit integer)
      • The basic reduction policy is to connect all return bits with all parameter bits that they depend upon ("reduction=basic")
        • An improved version (reduction=mincut) includes the minimal cut bits of the bit graph from the return to the parameter bits, assuming that the return bits have infinite weights, this is the default
      • properties
        • maxiter: maximum number of iterations for the coinduction, as every intermediate state is also valid
        • bot: default is basic only used for coind
        • mode: ind or coind, default is ind
        • reduction: reduction policy, either basic or mincut, default is mincut
        • dot: folder to output dot files for the bit graphs of the methods in different iterations and the call-graph
          • default: empty string, produces no dot files
        • csmaxrec: if > 0, each sub-analysis uses a call-string based handler for evaluation method invocations, using the computed summary edges as bot and the passed value as maxrec. Likely improves precision but also increases the size of the summary edges. Default is 0
  • the combobox labeled Min-Cut allows choosing between several minimum-cut algorithms for the analysis
  • the tabs below
    • Leakage: contains the leakage of information from higher or non-comparable security levels into each security level
    • Preprocessed: contains the SSA resolved version of the program with some operators replaced
    • Without loops: contains the program with all loops replaced by recursion
    • Output: Just some statistics
    • Node values: the values of the AST/PDG-nodes at the end of the analysis
    • Variable values: the values of the variables in the program at the end of the analysis
  • the graph view on the right
    • opens a new frame to view different graphs
    • the graphs are not created in the MIN output mode
    • zooming in the graph works by pressing the Shift key and moving the mouse up or down while pressing an arbitrary mouse key

License

It's licensed under the MIT license.

About

A quantitative information flow analysis on PDGs based on a data flow analysis

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages