Hello there. This text is not written by an AI. This is a repository with the code for the Story Beyond the Eye: Glyph Positions Break PDF Redaction paper.
Remediation efforts and questionable legality prevent me from releasing the code publicly to automatically break redactions: regardless of how spicy the results would be or how many sleepless nights were spent, I will not be caught giving a fully automatic machine gun to any small child.
Without further ado, let's talk about what this repository contains.
A submodule linking to a highly precise PDF Glyph Positioning, Text Analysis, and Parsing Library, PDF CMap Parsing Library. This compiles using just two makefiles. To build it:
# on some Ubuntu distributions (18.04), replace libfreetype-dev with libfreetype6-dev
sudo apt install cmake libfreetype-dev libfontconfig-dev libopenjp2-7-dev build-essential libjpeg-dev libnss-tls
cd PDF-Text-Analysis/lib/poppler
mkdir build
cd build
cmake ..
make
make install
# CMake configuration options can be set using the -D option. eg
# cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local -DCMAKE_BUILD_TYPE=release
cd ../../../adjustment
make VectorString.o
cd ..
make pts
and kablam, you are done.
You may need some other standard dependencies, but it should just work if you are familiar with Linux.
The primary method of using this is via the pts script, which supplies a
wealth of text positioning information to stdout by default. To test it out and
verify for yourself the promises of the paper are real, I've included a word
document from the washington post for you to look at (but not break).
In the parsing directory there are a number of Ruby scripts for translating
this UNIX-stype output to first-class objects. There is also the cmap script
which can retrieve character map information from PDFs. At the time of writing,
this was the only open source one available, though who knows what related work
has come out since then.
./PDF-Text-Analysis/pts inconspicuous_pdf.pdf 1
If you receive what(): locale::facet::_S_create_c_locale name not valid Aborted
then a UTF-8 locale has not been properly configured on your machine.
Uncomment the line "en_US.UTF-8 UTF-8" in /etc/locale.gen and then run the command
sudo locale-gen, which should fix the missing locale information. Other standard
unicode locales should also work (untested).
We have confirmed it works for a standard en_US (or similar) Unicode (UTF-8) locale:
LANG=en_US.UTF-8
LANGUAGE=
LC_CTYPE="en_US.UTF-8"
LC_NUMERIC="en_US.UTF-8"
LC_TIME="en_US.UTF-8"
LC_COLLATE="en_US.UTF-8"
LC_MONETARY="en_US.UTF-8"
LC_MESSAGES="en_US.UTF-8"
LC_PAPER="en_US.UTF-8"
LC_NAME="en_US.UTF-8"
LC_ADDRESS="en_US.UTF-8"
LC_TELEPHONE="en_US.UTF-8"
LC_MEASUREMENT="en_US.UTF-8"
LC_IDENTIFICATION="en_US.UTF-8"
LC_ALL=
- core: the wrapper for analyzing the PDF structure recorded in the modified poppler
- adjustment: a datatype for recording glyph displacements
- lib: the modified poppler
- parsing: ruby scripts for parsing glyph positioning information and cmap data
- utils: contains the actual analysis utilities (for this public version, just the print glyph information script)
I've included my scripts for locating redactions and fixing redactions under the defenses submodule. These are straightforward enough to run, although you may also need to download some dependencies. There are some test PDFs in this repository as well that you may try the code out on.
From the root of the repository, after building the PDF-Text-Analysis project:
pip3 install PyMuPDF
cd redaction-defenses
cp -r ../PDF-Text-Analysis .
For nonexcising redactions:
./trivial_redaction_loc.sh testcases/trivial_1.pdf
First, let me note I had to delete about 80%-90% of the code in this directory, and there are still some remnants of the early code for deredaction.
Thankfully, I don't believe what remains is too easy to adapt, and unfortunately a lot of it is tied up in the methods by which we correctly identify excising redactions.
First, install OpenJDK 17.0.6
sudo apt install openjdk-17-*
From the root directory of this repo:
pdfseparate inconspicuous_pdf.pdf inconspicuous_pdf.pdf-%d
cd excising
export PTS_PATH=<path to pts>
./location/cut_method.sh ../inconspicuous_pdf.pdf-11
There is also the possiblility of using
./location/ml ../../inconspicuous_pdf.pdf 10
This has a more general method of identifying redactions but relies on the old
c-src/core directory: to compile this, compile the lib/poppler directory like
before, cd into c-src/core and make.
Note this is not the full code to extract all the attack information used in
the paper. Think of it as an early version of that code with one leg cut off,
intended solely for the location of potentally vulnerable excising
redactions. The final attack did not use any of the shifts returned by ml,
but a later version of the algorithm did use the "BBox" tags.
if you get convert-im6.q16: attempt to perform an operation not allowed by the security policy PDF' @ error/constitute.c/IsCoderAuthorized/421` see
"https://stackoverflow.com/questions/52998331/imagemagick-security-policy-pdf-blocking-conversion"
Feel free to let me know if anything goes wrong. For now my email is
mb28@illinois.edu.
If you are a scholart, email me and I can give you the full code and datasets on valid scientific request, for reproduction purposes and to make your research easier!
There are also lots of small easter eggs buried around this repository that may help future research.
On (bad) advice, I was convinced to try out the ascii file, group, record, and unit separators:
034 28 1C FS (file separator) 134 92 5C \ '\\'
035 29 1D GS (group separator) 135 93 5D ]
036 30 1E RS (record separator) 136 94 5E ^
037 31 1F US (unit separator) 137 95 5F _
Output should still be readable, but you may need to translate these characters
by piping to the tr command. This is pretty straightforward I hope,
none-the-less.