Entries
Winning entries Winning authors Location of authors Bugs and (mis)features Fixing entries Updating author info Thanks for the help
Status
News Contest status Rules and guidelines Markdown guidelines Security policy
FAQ
Frequently Asked Questions How to enter Compiling entries Running entries How to help
About
Home page About the IOCCC The Judges Contact us
IOCCC image by Matt Zucker

The International Obfuscated C Code Contest

2011/dlowe - Most self deprecating

Neural network classifier

← 2011/borsanyi ↑ 2011 ↑ 2011/eastman → C code Makefile Inventory ✓

Author:

To build:

    make

Bugs and (Mis)features:

The current status of this entry is:

STATUS: missing or dead link or links - please provide it or them
STATUS: INABIAF - please DO NOT fix

For more detailed information see 2011/dlowe in bugs.html.

To use:

    ./dlowe -<n_iterations> corpus1/ [...] corpus0/ < start.net > trained.net

NOTE: In the above command, the directory args MUST end in a /.

Then to use trained.net:

    ./dlowe file [file ...] < trained.net

Try:

Try these pre-trained networks:

    ./try.sh

NOTE: The dlowe-aux-data directory, which try.sh refers to, was created by:

    tar -jxf dlowe-aux-data.tar.bz2

Judges’ remarks:

This entry is a ghoulish example of a brain (dead?) neural network classifier. It comes with a set of corpora for you to try.

ioccc-winlose-trained.net

You can create your own trained networks. For example, we trained a network on the C code of entries that were supplied to the 20Th IOCCC.

The ioccc-winlose-trained.net was trained using the earlystop.pl tool:

    rm -f ioccc-winlose-trained.net
    ./earlystop.pl ioccc-winlose-trained.net ioccc_won_training/ ioccc_lost_training/ ioccc_won_test/ ioccc_lost_test/

where:

NOTE: The above directories were not shipped and everything from the entries that did not win were flushed to /dev/null.

The earlystop.pl tool wrote:

    training 1 to 1001...
    training 1001 to 2001...
    prior error = 56.746974766043; current error = 58.095691135305

The resulting ioccc-winlose-trained.net file was included in the dlowe-aux-data.tar.bz2 tarball.

The dlowe.c was explicitly excluded from this trailing set, so this test is interesting:

    ./dlowe dlowe.c < dlowe-aux-data/ioccc-winlose-trained.net

NOTE: this is done in try.sh.

The match of dlowe.c to the lose/win trailed network was:

    dlowe.c 0.125108

NOTE: The non-artificially intelligent decisions of the IOCCC judges include random quirks and their non-linear judgment calls not found in the ioccc trained network. Just because C code is or isn’t scored highly by this network does not mean it will win or lose a future IOCCC. On the other hand … it just might!

other data sets to try

While we cannot supply you with the losing IOCCC source code, the author did supply a number data sets on which to test:

We created the english-trained.net as follows:

    ./dlowe -8000 dlowe-aux-data/english-1/ dlowe-aux-data/english-0/ < /dev/null > dlowe-aux-data/english-trained.net

We created the png-trained.net as follows:

    ./dlowe -8000 png-1/ png-0/ < /dev/null > dlowe-aux-data/png-trained.net

We created the xor-trained.net as follows:

    ./dlowe -8000 dlowe-aux-data/xor-1/ dlowe-aux-data/xor-0/ < /dev/null > dlowe-aux-data/xor-trained.net

NOTE: The directory args must end in a / for them to work.

Author’s remarks:

Synopsis

This is an artificially intelligent judging tool to help the IOCCC judges. Here’s to shorter, more frequent contests!

Description

This is a multilayer perceptron, a feedforward artificial neural network, which can be trained, using on-line backpropagation, to classify input files.

It has a fixed topology of 2^16 input neurons, 6 hidden neurons and 1 output neuron.

The neurons’ activation function is the logistic function 1 / (1 + e ^ -x).

Classifying

    ./dlowe file [file ...] < trained.net

To classify files, one specifies a trained network (on stdin) and one or more files to classify. The program will output one line per successfully-classified file to stderr, with the filename and the classification: a number between 0 and 1.

The interpretation of the classification number depends on how the network was trained, but it’s geared toward interpretation as a probability or a confidence.

Training

    ./dlowe -<n_iterations> corpus1/ [...] corpus0/ < start.net > end.net

NOTE: The directory args must end in a / for them to work.

To train a network, one specifies a starting network (on stdin), two or more corpora (directories containing training data), and the number of training iterations to run. The program will write some progress data to stderr and, when it’s done, will serialize the updated network to stdout.

If no input network is given, a random new network will be generated as a starting point (i.e. provide an empty stdin).

The first corpus will be assigned a target value of 1. The last will be assigned a target value of 0. Intervening directories (if any) will be assigned intermediate target values.

The learning rate is hard-coded as 0.3. No momentum factor is used.

png corpora

The dlowe-aux-data/png-1 corpus was obtained by manually scraping the first results from an https://images.google.com search for “obfuscate filetype:png”.

The dlowe-aux-data/png-0 corpus was obtained by manually scraping the first results from an https://images.google.com search for “obfuscate filetype:gif”.

Results:

produced a network with about 87% accuracy

English corpora

The english-1 corpus was obtained by manually scraping the first results from a https://www.google.com/webhp?lr=lang_en search for “paris filetype:txt”.

The english-0 corpus was obtained by manually scraping the first results from a https://www.google.com/webhp?lr=lang_fr search for “paris filetype:txt”.

Results:

produced a network with 100% accuracy.

xor corpora

The xor-1 corpus consists of two files containing 01 and 10 respectively.

The xor-0 corpus consists of two files containing 00 and 11 respectively.

Results:

Using these corpora as a training set takes about 8000 iterations to learn xor to within a tolerance of <0.01.

Limitations

The program can’t tell you anything meaningful about files with less than two bytes in them (I’m looking at you, smr.c!)

You must include the trailing directory separator on training directories (this allows the program to be portable without wasting precious bytes on figuring out how to concatenate directory and file names…).

Serialized network files are only portable between systems with the same floating-point representation and endianness.

Making sure not to overfit the network to the training data is a bit of a black art. I have enclosed earlystop.pl, a wrapper script that implements a simple ‘early stopping’ algorithm; other techniques are possible.

Bad input (e.g. nonexistent files, non-numeric number of iterations, etc.) tends to result in empty output.

Given exactly one corpus, the program will crash or produce garbage.

Leaks memory and file descriptors while processing files.

Will crash and die horribly if it runs out of memory.

The Microsoft C compiler doesn’t provide a dirent API, so to get this working on a Windows system you’ll need cygwin+gcc (tested) or a dirent compatibility library (untested, but they do exist).

Backpropagation doesn’t always converge: if you play with this long enough, you’ll eventually have a training session that completely fails to converge.

Obfuscation

… but mostly zombies!

Inventory for 2011/dlowe

Primary files

Secondary files

Jump to: top

Coda

Nu HTML check this web page
Jump to top Jump to Content Jump to Inventory