/Igor

Primary LanguageC

Igor: Crash Deduplication Through Root-Cause Clustering

Overview

Fuzzing has emerged as the most effective bug-finding technique. The output of a fuzzer is a set of proof-of-concept (PoC) test cases for all observed “unique” crashes. It costs developers substantial efforts to analyze each crashing test case. This, mostly manual, process has lead to the number of reported crashes out-pacing the number of bug fixes. Automatic crash deduplication techniques, which mostly rely on coverage profiles and stack hashes, are supposed to alleviate these pressures. However, these techniques both inflate actual bug counts and falsely conflate unrelated bugs. This hinders, rather than helps, developers, and calls for more accurate techniques.

Igor is a tool for automated crash grouping/deduplication. By minimizing each PoC’s execution trace, it can obtain pruned test cases that exercise the critical behavior necessary for triggering a bug. Then, Igor use a graph similarity comparison to cluster crashes based on the control-flow graph of the minimized execution traces, with each cluster mapping back to a single, unique root cause.

Igor helps a lot when you have many PoCs and would like to classify them into several groups according to the root cause, so that you don't need to analyze the PoCs one by one.

Here is a flow chart for overviewing the Igor's workflow.

More details about the project can be found at the paper.

Our presentation about Igor can be found at the video

Components

This repository is structured as follows:

  1. IgorFuzz (AFLplusplus): Our coverage decreasing fuzzer for test cases reduction.
  2. Smart_tracer (Pin): Our tracer to record control flow.
  3. Analyzer: Prune recorded execution traces and construct control flow graphs
  4. Cluster: Our cluster tool based on graph similar matrixs
  5. Evaluation: Our evaluation scripts used in Igor paper

IgorFuzz

We developed IgorFuzz based on AFLplusplus crash exploration mode. It can prune the paths that unnecessary for bug triggering very fast. Before using IgorFuzz, we suggest use afl-tmin to shrink the size of crash first, so that IgorFuzz will have better performance.

Installation and Usage

The installation and usage of IgorFuzz is completely same to the AFLplusplus' crash mode. Even time you want to launch IgorFuzz, you must confirm that you have put a PoC in input directory and set up output directory properly.

Reduction in parallel

IgorFuzz reduces one PoC at one time. To apply IgorFuzz on many PoCs parallelly, we provide users with mass_fuzz.sh. It will automatically run over and over again untill all PoCs in input dir are fuzzed.

Collect all PoCs you want to reduce in input directory(e.g., /home/my_pocs), and set up output dir(e.g., /home/trimmed/my_pocs).

The third arg is the number of PoCs you want to fuzz parallelly each time. The last arg is the duration the fuzzing last for(e.g., 1h2m3s).

Example:

$ ./mass_fuzz.sh /home/my_pocs /home/trimmed/my_pocs 30 10h

The form of result is: /home/trimmed/my_pocs/$the-name-of-a-PoC(like: id:000000,xxxxxxxx)/

mass_fuzz.sh renames fuzzed PoCs like: fzd_id:000000,xxxxx. So if there's something wrong with IgorFuzz or you want to shirnk all PoCs again, you can use ./clear_fzd.sh $INPUT_DIR to remove "fzd" prefix.

Tracing and Analyzing

To obtain precise execution traces (basic block level in default) of a specific binary, we need the following tools:

  • smart_tracer/calltrace_wrapper.py
  • analyzer/breakpoint_hit_counter.py
  • analyzer/find_crashing_addr.py
  • analyzer/trace_shrinker.py
  • analyzer/trace_pruner.py

For usages of the above tools, please check analyzer/README and smart_tracer/README.

Workflow

Execution traces need to be filted before constructing the control flow graph to be used to calculate the graph similarity. Follwing steps show how to do that.

STEP 1 - In the ASAN disabled environment

  • Using calltrace_wrapper.py to collect execution traces of the binary under test. Users can confiure which granularity they want to use, for now, we support instruction level, basic block level, and function call level.
  • Using trace_shrinker.py to filter out execution traces related to shared libraries.

STEP 2 - In the ASAN enabled environment

  • Using find_crashing_addr.py to find out the number of crashing addresses(the line number observed when the binary crashes). For each crashing address, repeat the following three steps:
    • Debug the binary under test, find the last function the binary calls before crashing, take down its caller's address(usually, the call instruction's address).
    • Using breakpoint_hit_counter.py to find out how many times the address mentioned above is hit before the binary crashes.
    • Copy the breakpoint hit count folder to our ASAN disabled environment.

STEP 3 - In the ASAN disabled environment

  • Debug the binary under test, find the same caller as the one in the ASAN enabled environment.
  • Using trace_pruner.py to prune redundant trace entries that are recorded after the point that the binary should have crashed. This step gives you a pruned traces directory for clustering.

Clustering

The cluster_utils folder contains the utilities for clustering.

cluster_utils/cluster.py will do everything for you, including construct control flow graphs based on pruned traces, calculate graph similarities and clustering. A simple example is listed below:

# '/traces/dir' is the pruned traces dir obtained by the "Tracing" work flow.
$ python3 cluster.py -i /traces/dir -o /output/dir

Ground-truth Benchmark

There are few public benchmark designed for the verification of crash grouping, especially for real world programs. In order to promote the research of crash grouping, we provide a a ground-truth benchmark for evaluating crash grouping techniques, containing 52 CVEs and more than 250,000 crashing test cases from 14 real world programs (generated over 58.7 CPU-years of fuzzing) for subsequent researchers, Igor also used this dataset to do the evaluations.

We are grateful to Magma and Moonlight for the original data and the methodology of establishing the ground truth data set. We used all their crashes and labels in the process of building the our benchmark, and further expanded the scale of the data set on their basis (more fuzzing time to generate more crashes).

Here are the links to our ground-truth benchmark:
benchmark
Every PoC is labeled with its root cause, user can get the label by parse the name of the PoC.

Building Magma targets approach can be found here
Building MoonLight targets approach can be found here

Contact

Questions? Concerns? Feel free to ping me via E-mail

TODO

  • Provide evaluation scripts we used in Igor paper
  • Provide link to Igor's dataset
  • Provide detailed tutorial for Igor system
  • Provide README for evaluation scripts
  • Provide scripts to do trace analyzing stuff automaticlly