A friendly request: Thanks for visiting control-flag GitHub repository! If you find control-flag useful, we would appreciate a note from you (to niranjan.hasab[email protected] or [email protected]). And, of course, we love testimonials!

— The ControlFlag Team

GitHub license

ControlFlag: A Self-supervised Idiosyncratic Pattern Detection System for Software Control Structures

ControlFlag is a self-supervised idiosyncratic pattern detection system that
learns typical patterns that occur in the control structures of high-level
programming languages, such as C/C++, by mining these patterns from open-source
repositories (on GitHub and other version control systems). It then applies
learned patterns to detect anomalous patterns in user’s code.

Brief technical description

ControlFlag’s pattern anomaly detection system can be used for various problems
such as typographical error detection, flagging a missing NULL check to
name a few. This PoC demonstrates ControlFlag’s application in the typographical
error detection.

Figure below shows ControlFlag’s two main phases: (1) pattern
mining phase, and (2) scanning for anomalous patterns phase. The pattern mining
phase is a “training phase” that mines typical patterns in the user-provided GitHub
repositories and then builds a decision-tree from the mined patterns. The scanning
phase, on the other hand, applies the mined patterns to flag anomalous
expressions in the user-specified target repositories.

ControlFlag design

More details can be found in our MAPS paper (https://arxiv.org/abs/2011.03616).

Directory structure (evolving)

  • src: Source code for ControlFlag for typographical error detection system
  • scripts: Scripts for pattern mining and scanning for anomalies
  • quick_start: Scripts to run quick start tests
  • github: Scripts and data for downloading GitHub repos.
    It also contains pre-processed training data containing patterns mined from
    6000 GitHub repositories using C as their primary language.
  • tests: unit tests


ControlFlag can be built on Linux and MacOS.


  • CMake 3.4.3 or above
  • C++17 compatible compiler
  • Tree-sitter parser (downloaded automatically as part of cmake)
  • GNU parallel (optional, if you want
    to generate your own training data)

Tested build configuration on Linux-based systems

  • CentOS-7.6/Ubuntu-20.04 with g++-v10.2.0 for x86_64

Tested build configuration on MacOS

  • MacOS Mojave v10.14.6 with clang-1001.0.46.4 (Apple LLVM version 10.0.1) for x86_64 (obtained from The Command Line Tools Package)


$ cd control-flag
$ cmake .
$ make -j
$ make test

All tests in make test should pass, but currently tests for Verilog are failing because of a version mismatch issue.
Verilog support is WIP.

Using ControlFlag

Quick start

Using patterns obtained from 6000 GitHub repos to scan repository of your choice

Download the training data for C language depending on the memory constraints of your device. Note, however, that using smaller datasets may lead to reduced accuracy in the results ControlFlag produces and possibly an increase in the number of false positives it generates.

Dataset name Size on disk Memory requirements Direct link gdown ID MD5 checksum
Small ~100MB ~400MB link 1gvUyRXq1SeZD9g3i__RaamYAMo_QaQIb 2825f209aba0430993f7a21e74d99889
Medium ~450MB ~1.3GB link 1zsCFJAKlZlSAWKPfBcVGcQNlFB5Gtwo3 aab2427edebe9ed4acab75c3c6227f24
Large ~9GB ~13GB link 1-jzs3zrKU541hwChaciXSk8zrnMN1mYc 1ba954d9716765d44917445d3abf8e85

$ python -m pip install gdown && gdown https://drive.google.com/uc?id=<id_from_table>
$ (optional) md5sum <tgz_file>
$ tar -zxf <tgz_file>

To scan C code of your choice, use below command:

$ scripts/scan_for_anomalies.sh -d <directory_to_be_scanned_for_anomalies> -t <training_data>.ts -o <output_directory_to_store_log_files>

Once the run is complete (which could take some time depending on your system and the
number of C programs in your repository,) refer to the section below to
understand scan output

Mining patterns from a small repo and applying them to another small repo

In this test, we will mine patterns from
Glb-director project of GitHub and
apply them to flag anomalies in GitHub’s brubeck project.

Simply run below command:

cd quick_start && ./test1_c.sh

If everything goes well, you can see output from the scanner in test1_scan_output
directory. Look for “Potential anomaly” label in it by grep "Potential anomaly" -C 5 \*.log, and you should see output like below:

thread_6.log-Level:TWO Expression:(parenthesized_expression (binary_expression ("==") (identifier) (non_terminal_expression))) found in training dataset:
Source file: brubeck/src/server.c:266:5:(s == sizeof(fdsi))
thread_6.log-Autocorrect search took 0.000 secs
thread_6.log:Potential anomaly
thread_6.log-Did you mean:(parenthesized_expression (binary_expression ("==") (identifier) (non_terminal_expression))) with editing cost:0 and occurrences: 1
thread_6.log-Did you mean:(parenthesized_expression (binary_expression ("==") (identifier) (null))) with editing cost:1 and occurrences: 25
thread_6.log-Did you mean:(parenthesized_expression (binary_expression ("==") (identifier) (identifier))) with editing cost:1 and occurrences: 5
thread_6.log-Did you mean:(parenthesized_expression (binary_expression (">=") (identifier) (non_terminal_expression))) with editing cost:1 and occurrences: 3
thread_6.log-Did you mean:(parenthesized_expression (binary_expression ("==") (non_terminal_expression) (non_terminal_expression))) with editing cost:1 and occurrences: 2

The anomaly is flagged for brubeck/src/server.c at line number 266.

Detailed steps

  1. Pattern Mining phase (if you want to generate training data yourself)

If you do not want to generate training data yourself, go to Evaluation step below.

In this phase, we mine the idiosyncratic patterns that appear in the control
structures of high-level language such as C. This PoC mines patterns from if
statements that appear in C programs.

If you want to use your own repository for mining patterns, jump to Step 1.2.

1.1 Downloading Top-100 GitHub repos for C language

Steps below show how to download Top-100 GitHub repos for C language
(c100.txt) and generate training data. training_repo_dir is a directory
where the command below will clone all the repos.

$ cd github
$ python download_repos.py -f c100.txt -o <training_repo_dir> -m clone -p 5

1.2 Mining patterns from downloaded repositories

You can use your own repository to mine for expressions by passing it in
place of <training_repo_dir>.

mine_patterns.sh script helps for this. It’s usage is as below:

Usage: ./mine_patterns.sh -d <directory_to_mine_patterns_from> -o <output_file_to_store_training_data>
[-n number_of_processes_to_use_for_mining]  (default: num_cpus_on_system)
[-l source_language_number] (default: 1 (C), supported: 1 (C), 2 (Verilog)

We use it as:

$ scripts/mine_patterns.sh -d <training_repo_dir> -o <training_data_file> -l 1

<training_dat_file> contains conditional expressions in C language that are
found in the specified GitHub repos and their AST (abstract syntax tree) representations.
You can view this file as a text file, if
you want.

Evaluation (or scanning for anomalies in C code from test repo)

We can run scan_for_anomalies.sh script to scan target directory of interest.
Its usage is as below.

Usage: ./scan_for_anomalies.sh -t <training_data> -d <directory_to_scan_for_anomalous_patterns>
 [-c max_cost_for_autocorrect]              (default: 2)
 [-n max_number_of_results_for_autocorrect] (default: 5)
 [-j number_of_scanning_threads]            (default: num_cpus_on_systems)
 [-o output_log_dir]                        (default: /tmp)
 [-l source_language_number]                (default: 1 (C), supported: 1 (C), 2 (Verilog))
 [-a anomaly_threshold]                     (default: 3.0)
scripts/scan_for_anomalies.sh -d <test_directory> -t <training_data_file> -o <output_log_dir>

As a part of scanning for anomalies, ControlFlag also suggests possible
corrections in case a conditional expression is flagged as an anomaly. 25 is the
max_cost for the correction — how close should the suggested correction be to
possibly mistyped expression. Increasing max_cost leads to suggesting more
corrections. If you feel that the number of reported anomalies is
high, consider reducing anomaly_threshold to 1.0 or less

Understanding scan output

Under output_log_dir you will find multiple log files corresponding to
the scan output from different scanner threads. Potential anomalies are reported
with “Potential anomaly” as a label. Command below will report log files
containing at least one anomaly.

$ grep "Potential anomaly" <output_log_dir>/thread_*.log

A sample anomaly report looks like below:

Level:<ONE or TWO> Expression: <AST_for_anomalous_expression>
Source file and line number: <C code with line number having the anomaly>
Potential anomaly
Did you mean ...

The text after “Did you mean” shows possible corrections to the anomalous expression.


View Github