CCDetector and BotnetDetectorComparer

Some days ago we finally made public two tools that were very important for starting this project. The tools are CCDetector and BotnetDetectorComparer. With these tools we created the experiments in the paper “An empirical comparison of botnet detection methods”. You can download them and use them to verify the paper and test more ideas. Please contact us if you need assistance.



A machine learning based detector of Command and Control channels in malware and botnet traffic. See please It uses an implementation of Markov Chains to model the state transitions of the traffic according to a model and detect similar behavioral traffic in other binetflow files.

It can also read binetflow files in real time from the network and print a nice ncurses interface with the states.



  • You should give it a labeled netflow text file with -f (generated by Argus) and use -r. See the example file for details. This will generate a MCModels folder full of the markov chain models for the tuples in the file.


After training the models with some file. Use -f to give a labeled binetflow file and use -e. Without any other option, a new file will be generated with the original binetflow information and an additional column with the predicted label based on the trained models. Also a sorted version of the binetflow file is created. No input is printed in the console.

Verification and performance metrics

To verify the results and know the performance metrics you should use another program called BotnetDetectorsComparer ( With this program you do: -f <binetflowfile>.labeled.sorted -t weight -T 300 -a 0.01

And find out the performance metrics according to a time window and weighted logic. Please see the papers.


  • 0.90 This is the first public version. Any problem please contact or


usage: ./ <options>
  -h, --help                 Show this help message and exit
  -V, --version              Output version information and exit
  -D, --debug                Debug level. E.g -D 3 .
  -f, --file                 Input netflow file to analize. If - is used, netflows are read from stdin. Remember to pass the header!
  -t, --time-threshold       First Threshold of time difference.
  -b, --bytes-threshold      First Threshold of bytes size.
  -d, --duration-threshold   First Threshold of duration.
  -T, --analyze-tuple        Analyze only this tuple and print detailed information for each netflow.
  -u, --tuple-mode           The tuple mode. can be 3 for sip-dip-dport or 4 for sip-sport-dip-dport
  -R, --thresholds           Threshold mode. Prints all the values of the features for training the thresholds.
  -r, --training             Training mode. Read one binetflow from -f file and outputs one Markov Chain for each label in the folder 'MCModels'. Don't use -r, -v or -e at the same time.
  -v, --validation           Validation mode. Read one or several binetflow files (comma separated in -f ) and consider them as training-validation. Use 10-folds to compute the models, applies the models and get the best thresholds. Don't use -r, -v or -e at the same time.
  -e, --testing              Testing mode. Read binetflows from -f file, Markov Chains from the folder 'MCModels', predicts for each tuple the chain (label) with more probability of generating it and it outputs a labeled netflow file. Don't use -r, -v or -e at the same time.
  -w, --without-colors       Do not use colors in the output
  -l, --state-length         Minimun length to consider the state string for analysis
  -p, --min-prob-threshold   Threshold to use when comparing each tuple to every model. If -q is specified then this is the minimal threshold to try. You can also specify it as 1e-10. The lower limit is 1e-45.
  -q, --max-prob-threshold   Maximum threshold to use when comparing each tuple to every model. -p must be specified also. You can also specify it as 1e-11. The lower limit is 1e-45.
  -L, --label                Print all the informatin about all the tuples with this label.
  -P, --print-mode           Print mode: normal, csv, oneline, and epoch. You can combine them with '-'.
  -a, --all-models           Generate and include all models in the process. By default it only uses the models of the C&Cs. Only for the training.
  -n, --num-folds            Number of folds in the k-fold validation.
  -s, --step-threshold       Step to use when moving the threshold. Defaults to 10. That is from 0.1 to 0.01. Use multiples of 10.


This is a program to compare different botnet/malware detectors based on network traffic. The idea is to read a netflow file that has a new column for each prediction of an algorithm and compare how each algorithm detects the traffic. It computes the FP, FN, TP and TN for each flow in a time window, by counting the errors per IP address. At the end of each time window several performance metrics are compared, and also at the end of the capture.


To use it you should give a binetflow file, the type of comparison and the width of the time window.

./ -f statisticGenerator.testcasewithheaders9.txt -t weight -T 300

Giving an alpha is also a good idea, if not the program will assume a default of 0.01 (like in our experiments)

With -p it will plot and open a window with the graph information for each method. With -P it will store the plots on disk. (format is in the help, but can by almost anything like png)

Any problem contact or


usage: ./ <options>
  -h, --help           Show this help message and exit
  -V, --version        Output version information and exit
  -v, --verbose        Output more information.
  -D, --debug          Debug. In debug mode the statistics run live.
  -f, --file           SORTED input netflow labeled file to analyze (Netflow or Argus).
  -t, --type           Type of comparison. Flow based (-t flow), time based (-t time), or weighted (-t weight). The weighted type is the new IP-based and time-based error metric.
  -T, --time           While using time based comparison, specify the time window to use in secodns. E.g. -T 120
  -p, --plot           Plot the fmeasures of all methods and show them on the display.
  -a, --alpha          In weight mode, use this alpha for computing the score (defaults to 0.4).
  -c, --cvs            Print the final scores in cvs format into the specified file. E.g. -c results.csv
  -o, --out            Store in a log file everything that is shown in the screen. -o all-info.txt
  -P, --plot-to-file   Instead of showing the plot on the screen, store it in a file. Type of plot given by the file extension. E.g. -P all-info.eps. The name of each algorithm is added at the beginning of the file name.

Written by Sebastian Garcia in Research Blog on Mon 12 October 2015.