Advanced Forensic Framework 4 (AFF4)
AFF4 was developed by Michael Cohen, Simson Garfinkel and Bradley Schatz. This page describes the basic design. See LibAFF4 for a description of how to use the sample implementation, library and tools.
Why did we want to design yet another forensic file format?
Traditional forensic file formats have a number of limitations which have been exposed over the years:
- Proprietary formats like EWF are difficult to implement and explain. EWF is a fairly complex file format. Most of the details are reverse engineered. Recovery from damaged EWF files is difficult as detailed knowledge of the file format is required.
- Simple file formats like dd are very large since they are uncompressed. They also dont store metadata, signatures or have cryptographic support.
- Traditional file formats are designed to store a single stream. Often in an investigation, however, multiple source of data need to be acquired (sometimes simultaneously) and stored in the same evidence volumes.
- Traditional file formats just deal with data - there is no attempt to build a universal evidence management system integrated within the file specification.
The previous AFF format made huge advancements in the field introducing excellent support for cryptography, digital signatures, compression and even the concepts of external referencing. It was time to gather up all the good things in AFF and redesign a new AFF4 specification.
We wanted to use a well recognized, widely supported and open bit level format. One of the strengths of AFF was the use of segments within the file format itself. It because obvious that the only requirement we have from an underlying storage mechanism is the ability to store blobs of data by name, and retrieve them by that name. How these are actually stored is quite irrelevant to us.
The sections below give a quick overview to some of the major ideas.
AFF4 is an object oriented architecture. We term the AFF4 universe the total set of objects which are known. Because AFF4 is designed to be scalable to huge evidence corpuses the AFF4 universe is infinite. All objects are addressable by their name which is unique in the universe. For example an AFF4 object might have a name of:
This is a standard URN notation object. The URN is unique. There will never be another object created anywhere in the universe with the same URN. Once objects are created their URN is fixed.
The AFF4 universe uses RDF to specify attributes about objects. In its simplest form (the one we use) RDF is just a set of statements about an object of the form:
Subject Attribute Value
******** Object urn:aff4:f3eba626-505a-4730-8216-1987853bc4d2 ***********
aff4:stored = urn:aff4:4bdbf8bc-d8a5-40cb-9af0-fd7e4d0e2c9e
aff4:type = image
aff4:interface = stream
aff4:timestamp = 0x49E9DEC3
aff4:chunk_size = 32k
aff4:compression = 8
aff4:chunks_in_segment = 2048
aff4:size = 10485760
This shows that the object named (the Subject) has all these attributes and their values. We call these relations or facts. The entire AFF4 universe is constructed around these facts. As we will see later facts can be signed by a person - which essentially has the person asserting that the facts are true.
AFF4 objects exist because they do something useful. What they do depends on the interface they present. Currently there are a few interfaces, the most important ones are the Volume interface and the Stream interface. An object's interface is a fact about the object with an attribute of aff4:interface. This tells us what the object can do for us.
On the other hand AFF4 objects can actually be different things and do what they do in a different way. The actual type of an object is specified by the attribute aff4:type. Whereas an interface tells us what the object can do for us, a type tells us what it actually is. (Its possible to change an object's type without changing its interface for example going from a ZipFile to a Directory volume. This does not affect any users of the object).
We define a Volume as a storage mechanism which can store a segment (bit of binary data) by name and retrieve it by name. Currently we have two volume implementations: a Directory and a ZipFile.
The Directory implementation stores the segments as flat files inside a regular directory on the filesystem. This is really useful if we want to image to a FAT filesystem since each segment is really small and we will not exceed the file size limitations. Its also possible to root the directory on a http url (i.e. the directory starts with http://somehost/url/). This allows us to use the image directly from the web - no need to download the whole thing.
Directory objects use FileLikeObjects (see below) to actually store the segments into different files. This means that Directory Volumes can be stored on HTTP or HTTPS servers, as well as regular directories.
The ZipFile implementation stores segments inside a zip archive. If the archive gets too large (over 4Gb) we use the Zip64 extensions to store offsets in 64 bits. This is nice since small volumes can just be opened with windows explorer. Its also really easy to extract the data out. A ZipFile volume uses a FileLikeObject to actually store the zip file.
This means that its possible to write a ZipFile volume directly onto a HTTP server and use the image directly from the server as well.
Example: http://www.pyflag.net/images/test.zip is an example of a small (about 1mb) AFF4 image.
Directory and ZipFile volumes can be easily converted from one to the other (i.e. unzip the ZipFile into a directory to create a Directory volume).
Streams are the basic interface for storing image data. Streams present a consistent interface which presents the methods of read, seek, tell' andclose. (Streams also supportwrite'', but thats a bit special because its how you actually create them).
As long as an AFF4 object presents a stream interface its possible to perform random reads within the body of data. Hence its possible to store any image data within the stream. The following section explain some of the specific implementations of streams.
The FileBacked object is a stream which stores data in an actual file on
the filesystem. The location of the file is determined from the file's
URN. Since a URN is a superset of URLs, URLs are also valid URNs. This
means that something like
HTTP is ubiquitous and easy to deploy. Since URLs are also valid URNs, its possible to specify that an AFF4 volume be stored or read from a HTTP server. This implementation uses the Range HTTP header to read specific byte ranges from the server - so network traffic between the client and server is minimal. Its possible to examine a remote image over HTTP without needing to copy the whole thing down.
This is excellent when you just want to have a quick look at a remote image without needing to download the whole thing.
For security reasons its recommended write support be restricted in some way (e.g. passwords, SSL certificates etc). Read support can be provided freely if the volume is encrypted. Securing the web server is outside the scope of AFF4.
Segments are components stored directly within the Volume. Recall that a volume is simply an object which stores and retrieves segments. Segments also present the stream interface, but practically they should generally be used for smaller streams because it may be expensive to seek within compressed segments.
Segments are particularly useful when you dont have an imaging tool handy and you want to create a logical image of a subset of a filesystem (that is you want to image some files from a filesystem rather than a forensic image of the filesystem itself). This could happen if you can not take the server down for incident response or if the filesystem is just so big and you know most of it will not be relevant.
In that case there is nothing simpler than just to open up windows explorer - right click and send to a compressed folder. A regular zip file is also an AFF4 volume!!! The files within it are stream objects and libaff4 will recognize them as such. Larger segments can be converted to Image streams later (and signed, encrypted etc).
Although segments are great for small files, for very large images we cant really use those because we could not compress them efficiently. Therefore we have an image stream.
The Image stream stores the image in chunks. Each chunk (typically 32kb) is compresses and a group of chunks (called bevies) are stored back to back inside a bevy segment. Segments are named according to the scheme: URN_OF_IMAGE_STREAM/0000000, URN_OF_IMAGE_STREAM/0000001 etc.
The offset of each chunk within the bevy is stored in an index segment (with a. idx extension). Here is an example:
Here is a short python program to unpack an Image stream:
outfd = open(OUTPUT_FILE,"w")
count = 0
idx_segment = volume.read(STREAM+"/%08d.idx" % count)
bevy = volume.read(STREAM+"/%08d" % count)
indexes = struct.unpack("<" + "L" *
for i in range(len(indexes)-1):
chunk = bevy[indexes[i]:indexes[i+1]]
count += 1
A lot of work in digital forensics involves copying data around. For example, carving files usually results with the carved files copied out of the image for testing. If you image a RAID array separately you end up with 3-5 disk images and typically you will need to copy them into a logical image (unless your favourite software supports RAID reconstruction). When you copy a file out of the image using sleuthkit, you are actually copying bits of data directly from the image.
All these copies are wasteful of disk space. They are also hard to manage because pretty soon you end up with lots of copies of the same data in different ways. There must be a better way!!!
Now there is. By having the underlying forensic format doing all the mapping its possible to use tools which are not capable of doing these transformations themselves. This is all about tool reuse. For example suppose you have a carver which is used to work on dd images. But you want to use it on the virtual memory image of the firefox process. In the past you had to copy the virtual memory out (it could be 2-4gb) then run the carver on it, and possibly end up with about 3 or 4 copies of the same data - for each process address space!!!
Its much easier to have volatility create the initial maps for each process (with zero storage overheads), and then carvers can just use the maps without understanding anything about memory forensics. In this way the AFF4 format is more of an interchange format - allowing tools to be used on the results from other tools.
The map stream is an AFF4 object which contains a segment called map. Here is an example:
This map was generated by sleuthkit for an ext2 file. The first number of the offset in the image. The second number is the offset in the file and the third number of the URN of the object to read from (the target). The above map basically says that the byte range from 0-12288 in the file should be read from aff4:f3eba626-505a-4730-8216-1987853bc4d2 offset 1601536 to 1614848, etc.
Using the fsbuild tool its possible to create map streams for all files in an image at virtually zero storage requirements. Then its possible to use other tools which may not know how to read filesystems to examine the files.
Following is an example of a 3 disk RAID system which was acquired with into three separate streams:
To make this work we need to tell AFF4 that the map should be repeated. We do this by setting attributes on the map objects:
- Extending the Advanced Forensic Format to accommodate Multiple Data Sources, Logical Evidence, Arbitrary Information and Forensic Workflow, by M. I. Cohen, Simson Garfinkel and Bradley Schatz, DFRWS 2009, Montreal, Canada.
- Wirespeed: Extending the AFF4 forensic container format for scalable acquisition and live analysis, by Bradley L. Schatz, DFRWS US 2015