Fuzzy Hashing and E-Discovery

Recent work has made me consider an interesting role fuzzy hashes could play in E-Discovery.

In the last year I've worked a few intellectual property theft cases where Company A has sued Company B claiming Company B stole IP from Company A in the form of documents, design drawings, spreadsheets, contracts, etc.

In these cases Company A has requested that Company B turn over all documents that may pertain to Company A or Company A's work product, etc. with specific search terms provided and so on.

Company B argues they can't comply with Company A's request because they have documents relating to Company A and Company A's work product as a result of market research for the purposes of strategic planning and that turning over all of those documents would damage Company B.

In such cases, if Company A is concerned that Company B has stolen specific documents, maybe a better approach would be to request that Company B run ssdeep or another fuzzy hashing tool against all of their documents and turn over the fuzzy hashes.

Company A can then review the fuzzy hash results from Company B without knowing anything about the documents those hashes came from. They can compare the set of hashes provided by Company B against the set of fuzzy hashes generated from their own documents and make an argument to the judge to compel Company B to turn over those documents that match beyond a certain threshold.

24:DZL3MxMsqTzquAxQ+BP/te7hMHg9iGCTMyzGVmZWImQjXIvTvT/X7FJf8XLVw:J3oy+x/te7qmNmlYvX/xp8W

Facebook Artifact Parser

If you have a Facebook account, take a look under the hood some time by viewing the source in your browser while you're logged in. Imagine having to deal with all of that for a digital forensics investigation. It's mind numbing, especially if all you want is who said what and when. I spent the better part of today brushing up on Python's regular expression implementation and put together this Facebook Artifact Parser that does a decent job of parsing through Facebook artifacts found on disk (as of the time of this writing).

In my case, I made use of this by first recovering several MB worth of Facebook artifacts from disk and I combined all of these elements into one file. Having done that, run this script from the command line giving the name of the file as the only argument. It works on multiple files as well.

Yahoo! Messenger Decoder Updated

I'm working yet another case that involves Yahoo! Messenger Archives. I tried using JAD Software's excellent Internet Evidence Finder for this and it worked pretty well, but in the interest of double-checking my tools, I brushed off my old yahoo_msg_decoder.py script that I'd written a few years ago. It used to be interactive, meaning it was run with no arguments and would prompt for a username and a filename to parse, this was less than ideal for parsing a large number of files.

I have remedied that situation. The script now takes three arguments, one optional. The first is the username for the archive. Yahoo! Messenger Archives are xor'd with the username. The second argument is the name of the other party to the conversation and the third argument is the name of the dat file to process.

The nice thing about this is that you can now create a for loop like the following from a Linux environment and parse multiple files at once:

for i in $(ls *.dat); do echo; echo "== Parsing $i =="; yahoo_msg_decoder.py --username=joebob --other_party=billybob --file=$i; echo "== Finished parsing $i =="; echo; done

The output of this for loop can be redirected to a file.

My script is still not perfect. On some dat files it doesn't properly xor the data and yields garbage. I have not determined why that is the case yet.

As for IEF, I'm not sure why, but running it over the same dat files as my script, it dropped some portions of the conversation. I will be reporting the issue to JAD. But it's yet another reminder of the importance of testing your tools and confirming results.

update: After posting this, I remembered that Jeff Bryner had written a utility for this and it is still vastly superior to my own. I just verified that the link I have to his yim2text still works. Check it out.

Awk regtime bodyfile adjustment

Here's an awk one liner for adjusting regtime bodyfile time stamps, in this case we're adding 600 seconds:

awk -F'|' 'BEGIN {OFS="|"} {$9=$9+600;print}'

One thing to consider when adjusting time stamps to compensate for clock drift, clocks don't drift all at once, but over days, weeks and months. Adjusting time skews affects everything all at once.

Time again

I gave a version of the Time Line Analysis talk at Cyber Guardian earlier this week. Some in the room asked if the slides would be made available. As promised, here is a link to the deck. Enjoy.

MapReduce for simpletons

Data reduction redux and map-reduce is the title of my latest post at the SANS Digital Forensics Blog. I mentioned in my previous post there on using least frequency of occurrence in string searching that there would be a follow up.

The point of the new post is to sing the praises of @strcpy over on the Twitters. He helped me out by writing a short shell script that is, in essence, map-reduce for simpletons like me. I am constantly amazed by some of the members of the info sec community who will take time to help out near total strangers.

strcpy's script wasn't just helpful, it was educational. I'd read about map-reduce before, but it never really clicked until I saw strcpy's script. The scales have fallen from my eyes and I'm now adapting his script for other kinds of tasks.

Check out the post and if you find it beneficial and you ever get to meet strcpy in person, buy him a drink or a meal and tell him thanks, I plan to do the same one day.

Scalpel and Foremost

The crew over at Digital Forensics Solutions announced the release of a new version of Scalpel with some exciting new features. Check out their post for the full details, but here are three I was most interested in:

Parallel architecture to take full advantage of multicore processors

Beta support for NVIDIA CUDA-based GPU acceleration of header / footer searches

An asynchronous IO architecture for significantly faster IO throughput

Digital forensics is time consuming so any speed gains we can make are welcome ones.

Over the last few days, I've had a chance to play with the new version of scalpel on my 64-bit Ubuntu system with 7GB of RAM. I downloaded the source and followed the directions in the readme to configure and compile the binary.

I then ran some carves against a 103GB disk image from a recent case. The command line I used was:

scalpel -b -c /etc/scalpel.conf -o scalpel-out/ -q 4096 sda1.dd

The -q option is similar to foremost's -q option in that it tells scalpel to only scan the start of each cluster boundary for header values that match those specified in the config file. In my test, I used the two doc file signatures in the supplied example scalpel config file. The nice thing about -q is that you can provide the cluster size. With foremost -q will scan the start of each sector by default, you'll have to also add -b to get similar functionality out of foremost.

I ran scalpel with the Linux time command so I could determine how long the command took to complete. Scalpel carved 6464 items that had byte signatures matching those in the configuration file. According to the time command, this took 52 minutes and 40 seconds.

Manually verifying that all 6464 files are Word docs would be time consuming. In lieu of that, I followed Andrew Case's suggestion and used the following command from within the scalpel-out directory:

for i in $(find . | grep doc$); do file $i; done | grep -i corrupt | wc -l

The result was that 2707 of the 6464 files were found to be "corrupt" according to the file command. This is not an exact measure of the accuracy of scalpel's work, but it gives us a ballpark figure. If my math is correct that's a false positive rate of 41%. Just remember, these are rough figures, not exactly scientific.

Next I configured foremost to use the exact same configuration file options and similar command line arguments (recall I had to use -b with foremost) and ran the carve against the same image. The command line I used was:

foremost -c /etc/foremost.conf -i sda1.dd -o foremost-out -q -b 4096

Again, I used the time command to measure how long this took, 47 minutes and 32 seconds later, foremost finished having carved 6464 files. I used the same measure for accuracy as with scalpel, running the following command from within the foremost-out directory:

for i in $(find . | grep doc$); do file $i; done | grep -i corrupt | wc -l

The result was that 2743 files came back as "corrupt" according to the file command. Interesting. Both tools used the exact same signatures, both carved exactly the same number of files, yet foremost was approximately 1% less accurate, though at 47 minutes compared to scalpel's 52 minutes, it was almost 10% faster.

Conclusion:
It's hard to draw conclusions from one simple test. I think it's great that scalpel is under active development and for those who can take advantage of the CUDA support, it could be a huge win in terms of time and time is against us these days in the digital forensics world.

The other big plus, is that it's great to have another tool that we can use to test the results of another tool. I will continue to experiment with scalpel and look forward to future developments and I thank the developers of both tools for their contributions to the community.