You can transform the PDF into uncompressed form using
pdftk. Most metadata will then be immediately visible (and removable, provided you repair the file with pdftk afterwards). Same goes for "non-immediately-PDF" code (you can see that with tools such as PDFid). You can for example easily alter the trailer where fields such as /ID are to be found:
/Info 104 0 R
/ID [<81b14aafa313db63dbd6f981e49f94> <81b14aafa313db63dbd6f981e49f94>]
(in the case of ID, just replace with a sequence of random hex of the same length. No repair necessary. pdftk compress command is advised to save disk space).
It much depends on what kind of redaction you're trying to achieve. "Semi-inadvertent" metadata such as the ID (above) can be easily removed either directly or by re-saving the PDF removing unused objects and previous revisions of extant objects, which could contain sensitive information either intentionally or unintentionally:
7.5.6 Incremental Updates
The contents of a PDF file can be updated incrementally without rewriting
the entire file. When updating a PDF file incrementally, changes shall be appended
to the end of the file, leaving its original contents intact
Some filters may contain garbage without this compromising PDF legibility (for example the DCTDecode filter used to store a JPEG entity within the document. The filter supplies its own data size field internally; so what happens is that the "outer" PDF reader gathers the DCTDecode-encoded object and passes it to the filter, which gives back a raster image. The original object might contain extra bytes that the filter will ignore and discard, and which may be meaningless, or may contain pieces of random memory from the encoding computer (and possibly sensitive information), or may contain the same intentionally. Actually, being a JPEG, a DCTDecode object could contain EXIF information and this might be sensitive (e.g. GPS positioning and the like).
For example if I take an image containing a copyright
$ exiftool istockphoto_2425717-getting-a-call.jpg | grep Copyright
Profile Copyright : Copyright (c) 1998 Hewlett-Packard Company
and convert it to PDF
$ convert istockphoto_2425717-getting-a-call.jpg test.jpg
...the original string is still there:
$ strings test.pdf | grep Copyright
Copyright (c) 1998 Hewlett-Packard Company
To remove it, I would need to extract the JPEG(s), remove the EXIF and other data tags from them, and then re-embed them back. It could conceivably be done using e.g. iText, or a state machine identifying the streams
/ColorSpace 8 0 R
ÿØÿà^@^PJFIF<this is the JPEG file...>
saving the stream object as a JPEG, manipulating it, and writing it back. Now the stream length needs to be corrected, and the index offsets will be changed, yielding a broken PDF; pdftk can reconstruct the index from this broken PDF generating a clean one.
Yet other kinds of metadata, designed for stealth, may not be so easy to detect, much less remove. The possibilities there are endless. Some of the sneakier:
- PDF has an internal index, so the actual ordering of most object entities inside the file does not matter. Find a metric - lexicographic could do - to define a "natural" order of those objects. If you have N objects, you have N! possible shuffles of those objects, and can then encode log2(N!) bits of information in the way they are ordered. Remove with: a tool capable of reordering the objects in the PDF file.
- Fonts can map arbitrary symbols to glyphs and glyphs are ordered within a font. Again you can choose the mapping so it carries information. Remove with: most PDF "font optimizers".
- Several objects allow positioning at much higher precision than actually needed for typesetting. Assuming there's no typesetting difference between the coordinate values 3.1415 and 3.1416, a character positioned at 3.1415064 could "encode" the ASCII code 64. Remove with: some tool (iText) capable of extracting and processing every entity in the PDF and align it to a grid with arbitrary precision (here 0.0002), thereby rounding 3.1415... to 3.1416.
A not always optimal, but working, way of doing all the above is to regenerate the PDF from scratch by e.g. printing it to another PDF device after typesetting it in a different format. A radicalization of this approach is to typeset the PDF onto a raster canvas and generate a PDF out of that. The resulting PDF will be far bulkier since it now basically consists of a series of TIFF or JPEG images, one per page. It will also be, of course, non-searchable.
It is still possible to steganograph information that will survive such treatment by using visible character kerning encoding, that will be replicated on the rasterized version. Some "safe" documents are typeset that way - if the document leaks, accessing a copy of it will reveal the identity of the responsible party; two copies assigned to different personnel, superimposed against a sunny window, will show the characters aren't exactly aligned in the two copies. A scheme used to do the trick is called QIM (other schemes exist).
Depending on the degree of control the encoding party has on the document, you might therefore never be sure you've successfully "anonymized" a document. For sufficiently large values of paranoid, even going all the PDF2ASCII way might not touch information encoded in typos or, who knows, synonym choice (I think I still have some texts that demonstrate such an encoding, but unfortunately they're in Italian :-) ).