Monday, June 22, 2020

Comparative analysis between Bindiff and Diaphora - Patched Smokeloader Study Case

This article presents a comparative study case of diffing binaries using two technologies: Bindiff [1] and Diaphora [2]. We approached this topic in a Malware Analysis perspective by analyzing a (guess which malware family?) Smokeloader (! :D) campaign.

In August 2019, I spotted this campaign using patched samples of Smokeloader 2018 samples. This specific actor patched binaries to add new controllers URLs without needing to pay extra money (Smokeloader's seller charges extra-fee for C2 URL updates). This campaign was described in more detail in this previous article [3].

More details about the original samples [4][5] analyzed in this article can be found in the following tables:

 Filename:                        smokeloader_2018_unpatched.bin 
 Size:
 33792 Bytes
 File type: PE32 executable (GUI) Intel 80386, for Microsoft Windows
 md5: 76d9c9d7a779005f6caeaa72dbdde445
 sha1: 34efc6312c7bff374563b1e429e2e29b5da119c2
 sha256: b61991e6b19229de40323d7e15e1b710a9e7f5fafe5d0ebdfc08918e373967d3  

 Filename:                        smokeloader_2018_patched.bin
 Size: 1202732 Bytes
 File type: PE32 executable (GUI) Intel 80386, for Microsoft Windows 
 md5: 7ba7a0d8d3e09be16291d5e7f37dcadb
 sha1: 933d532332c9d3c2e41f8871768e0b1c08aaed0c
 sha256: 6632e26a6970d8269a9d36594c07bc87d266d898bc7f99198ed081d9ff183b3f  

The following tables hold details about the unpacked code dumped from "explorer.exe" used in this article [6][7]. 

 Filename:                          explorer.exe.7e8e32c0.0x02ee0000-0x02ef3fff.dmp 
 Size: 81920 Bytes
 File type: data
 md5: 711c02bec678b9dace09bed151d4cedd
 sha1: 84d6b468fed7dd7a40a1eeba8bdc025e05538f3c
 sha256: 865c18d1dd13eaa77fabf2e03610e8eb405e2baa39bf68906d856af946e5ffe1  

 Filename:                       explorer.exe.7e8df030.0x00be0000-0x00bf3fff_patched.dmp
 Size: 81920 Bytes (yes, same size)
 File type: data
 md5: d8f23c399f8de9490e808d71d00763ef
 sha1: e1daad6cb696966c5ced8b7d6a2425ff249bf227
 sha256: 421482d292700639c27025db06a858aafee24d89737410571faf40d8dcb53288  

Summarizing the main changes implemented by this patch are: 
  • wipes out the code for decrypting C2 URLs;
  • replaces it with NOPs and hardcoded C2 URL string; and
  • preserves the original size of decryption function to not disrupt offsets;
Figure 01 and 02 presents the graph of the original code. Figure 01 is the code used for indexing a table of encrypted C2 URLs payloads. Figure 02 lists the code used for decrypting the C2 URLs. This function is called in other parts of the code (not only by the function shown in Figure 01) - this is why they are not merged in one function. We labeled functions (e.g. "__decrypt_C2_url" and "__decrypt_c2_algorithm") in this assembly code to make it easier to read. 

Figure 01 - Original code used for decrypting C2 URLs.

Figure 02 - Original encryption scheme used for decrypting C2 URLs.

Figure 03 and 04 shows the same functions on the patched version of Smokeloader. We can notice that the first function is the same as the unpatched version but the second function was replaced. This second function returns the address for the hardcoded URL in ECX. More details about how it works can be found in this article [3].

Figure 03 - Patched code used for decrypting C2 URLs.

Figure 04 - Patched code returning decrypted C2 URL string.

The next sections describe the output of Diaphora and Bindiff when diffing the samples above. 

.::[ Diffing using Diaphora ]

Diaphora is an Open Source binary diffing tool that uses SQLite as an intermediate representation for storing code and characteristics of reversed binaries [8]. It implements many diffing heuristics (strategies) directly on top of this database. The main advantage of this approach is that Diaphora is technology agnostic - this means that it does not depend on any reversing framework such as Ghidra, IDApro, or Binary Ninja

It can even compare reversing databases built up using one specific tool with projects using another tool. This characteristic facilitates collaboration among researchers. Another big advantage is that by using SQLite for describing its heuristics it makes the processing of adding a new heuristic as "simple" as writing a new SQL so more people can contribute to the growth of the project and more experimental heuristics can be quickly prototyped and verified. 

In the experiment described in this section, we used Diaphora version 2.0.2 (released in October 2019) and IDApro 7.5. Diaphora works as an IDApro Python script and can be easily executed by "File -> Script File" or "alt + F7". Figure 05 shows its main interface. 

Figure 05 - Diaphora main Dialog Interface

This interface is a little bit confusing at a first sight especially for users that did not go through the documentation before trying to use it. It expects the user to input both SQLite databases to be compared, boundaries (for the working database), and set up some checkboxes with options. 

First, we used Diaphora over the reference database (unpatched Smokeloader) for extracting characteristics and generating our reference Diaphora SQLite database. For doing this we just need to open the reference database on IDA, open Diaphora, and fill the "Export IDA database to SQLite" input field. Diaphora will export its SQLite database to the same base directory of the IDA working files. 

After executing Diaphora on our patched Smokeloader using out saved labeled database of Smokeloader 2018 as a reference in Diaphora we get four new tabs in IDA
  • Best Matches - common functions to both databases. The ones with 100% match ratio;
  • Partial Matches - all functions that are not Best matches and not unmatched; 
  • Unmatched in Primary - all functions in the first database that are not present in the second;
  • Unmatched in Secondary - all functions in the second database that are not present in the first;
Figure 06 shows the content of the "Best Matches" tab in our example. 

Figure 06 - Diaphora Best Matches tab

Each row shows function labels in primary and second databases, matching ratio (which goes from 0 to 1), amount of basic blocks in each database and information about which heuristic was used to compare both functions. Diaphora provides features to importing features (such as commends and function labels) from the reference database to the target database. Usually, you will want to copy all annotations from one database to another in this "Best Matches" tab. By checking this tab we can see that few core functions are kept intact in the patched version and we are facing two very similar applications. 

Figure 07 presents the "Partial Matches" functions.

Figure 07 - Partial Matches functions and our "__decrypt_C2_url" function right there with 0.978 matching ratio.

Diaphora provides very fine level diffing and most of these functions are basically unmatching constants. These matches are marked in yellow in the graph diff view. Figure 08 and 09 shows the "__set_file_attributes" function and its match in the primary database. We can clearly see that they actually are the same function. 

Figure 08 - Diffing "__set_file_attributes" function assembly view.

Figure 09 - Diffing "__set_file_attributes" function using graph view.

Figure 10 shows the relevant patch we are interested in the "__decrypt_C2_url" function. As we can see both functions are basically identical until the jump instruction. The function jumps to what I labeled as "__decrypt_C2_algorithm" and is used as a function in other places around the code. 

Figure 10 - "__decrypt_C2_url" graph diff pointed in the "Partial Matches" tab

What disappointed me a little bit was that Diaphora did not include the rest of the code of "__decrypt_C2_algorithm" in this function.  This move bumped up the matching ratio and this was misleading when prioritizing what to manually analyze. The "__decrypt_C2_algorithm" function shows up in the "Unmatched in Secondary" tab. This is a good thing as functions in this tab should be the priority in this kind of analysis. In our example, we got 18 functions (out of 52) to analyze marked in the "Unmatched in Secondary" tab. Figure 11 shows this tab and Figure 12 shows the graph view of this function.

Figure 11 - "Unmatched in Secondary" tab and patched "__decrypted_C2_algorithm" function

Figure 12 - Patched version of "__decrypt_C2_algorithm" function

The nicest thing about using Diaphora is that all this analysis could be easily shared by sharing compressed SQLite databases around. So that is it for the Diaphora side. 

.:: [ Diffing using BinDiff ]

BinDiff is an executable-comparison tool created by Zynamics [9] (called SABRE in earlier days) in 2007. Zynamics was acquired by Google in 2011 and BinDiff became freeware in 2016 [10][11]. BinDiff is a plugin of IDAPro and works directly over IDBs (IDA pro Database). Because of this design, BinDiff requires users to have an IDA license (#notcool). BinDiff 6.0 supports also Ghidra using this extra plugin called BinExport [12] which implements something similar to Diaphora's design.

OALabs released a very didactic video tutorial on BinDiff [13]. This video also teaches how to install BinDiff. We used BinDiff 6.0 and IDAPro 7.5 in this experiment. BinDiff adds a new option to the "File" menu in IDA and to compare two executable is as easy as opening a new IDB on top of the current one. Figure 13 shows the BinDiff option in IDA

Figure 13 - BinDiff option in IDAPro

After loading an annotated IDA database using BinDiff, it will add 4 new tabs:
  • Primary Unmatched - functions in the primary database that did not match any function in the secondary database;
  • Secondary Unmatched - functions in the secondary database that did not match any function in the primary database;
  • Statistics - general similarity information about both executables;
  • Matched Functions - all functions with matches and their respective similarity index. 
The two first tabs hold the same information as their correlated tabs in Diaphora. Statistics tab provides high-level information about the matching process. Information in this tab can be used for quick knowing if the binary is a variation of the reference database. Figure 13 shows the data presented in the Statistics tab after loading our reference Smokeloader database against the patched one using BinDiff

Figure 14 - Statistics tab and confidence and similarity indexes

BinDiff calculated a similarity index of 95% (with 99% confidence). This means that we are likely dealing with two versions of the same software. The other metrics are more about counters and general information about both databases. 

Figure 14 shows the Matched Functions tab and the similarity index for each match. It is also possible to see the heuristic used for each match. 

Figure 15 - Matched Functions tab

BinDiff managed to match 51 out of 52 functions. This is a very good result. Besides that Bindiff managed to find out the two most affected functions by the patch: (i) "__decrypt_C2_algorithm" and (ii) "__decrypt_C2_url". It is possible to zoom in and visualize changes in graphs of the function matched. 

Figure 16 shows changes in function "__decrypt_C2_url" (73% of similarity and 97% of confidence).

Figure 16 - Changes in function "__decrypt_C2_url"

As we can see, BinDiff hits the bullseye and detects exactly the changes without splitting the function into two parts. The way BinDiff organizes its graphs makes changes really easy to visualize. 

Bindiff also matches the "__decrypt_C2_algorithm" function with some other function located at "0x00BE19A5" using the "loop count" heuristic but the similarity index is only 19% and confidence is 27%. This means that Bindiff matched the "__decrypted_C2_algorithm" function, which was wiped out with the patch, with some other random function. I don't even consider this a false-positive as results clearly state that this match has low confidence. It is useful to get a spotlight pointed to this function - this for sure is a good place to start an analysis in this specific scenario. 

.:: [ Conclusions ]

For the specific malware analysis problem discussed in this article, we definitely got better results using BinDiff than Diaphora. I also feel that all fine program analysis used in heuristics applied by BinDiff makes a big difference in the final result. 

In terms of general design (not taking in consideration heuristics), I think Diaphora has more advantages than Bindiff, because of its intermediate representation using an open specification format as SQLite and SQL for modeling heuristics. Diaphora is also Open Source so there is a task force sustained by a community in order to improve heuristics and this is the way to go IMHO

Diaphora also takes boundaries as parameters and this can be very useful in case of analyzing big databases especially when analyzing patches for vulnerability development purposes. In line with that, this article focuses on malware analysis but these same technologies could also be used for analyzing vulnerabilities and its patches - ammunition to another blog post. 

.:: [ References ] 

[1] https://www.zynamics.com/bindiff.html
[2] http://diaphora.re/
[3] http://security.neurolabs.club/2019/08/smokeloaders-hardcoded-domains-sneaky.html
[4] https://www.virustotal.com/gui/file/b61991e6b19229de40323d7e15e1b710a9e7f5fafe5d0ebdfc08918e373967d3
[5] https://www.virustotal.com/gui/file/6632e26a6970d8269a9d36594c07bc87d266d898bc7f99198ed081d9ff183b3f
[6] https://www.virustotal.com/gui/file/865c18d1dd13eaa77fabf2e03610e8eb405e2baa39bf68906d856af946e5ffe1
[7] https://www.virustotal.com/gui/file/421482d292700639c27025db06a858aafee24d89737410571faf40d8dcb53288
[8] https://www.youtube.com/watch?v=eAVfRxp99DM (BSides Joxean)
[9] https://www.zynamics.com/company.html
[10] https://security.googleblog.com/2016/03/bindiff-now-available-for-free.html
[11] https://www.zynamics.com/bindiff/manual/#N20140
[12] https://github.com/google/binexport/tree/v11/java/BinExport
[13] https://www.youtube.com/watch?v=BLBjcZe-C3I (BinDiff OALabs)


No comments:

Post a Comment