Project is not ready, avoid opening issues

ML-IDS

Applying machine learning principles to the information security field through intelligent intrusion detection systems.

This is a three phases project that can roughly described as

• Data gathering
• Model developpement
• Hardware implementation

Project structure

This project has four folders

• data -> Dataset(s) related to the project
• model -> Three machine learning model that I used to get my result (structure only) + a DataManager class to turn the dataset into learnable features (KDD Cup dataset only)
• implementation -> Empty for now, will contain the actual C++ implementation that could be deployed on the proposed x86-64 or ARM machines.
• papers ->

Data gathering

There are a few freely available datasets that applies to the field, I could attempt to create my own set by setting up a virtual machine exposed to the Internet but it is unlikely that I will record any privilege escalation attack (or any sophisticated attacks really) only DoS, DDoS, SSH bruteforce and scans could really be recorded.

Freely available datasets

• DARPA Intrusion Detection Data Sets (1998-2000)
• KDD Cup (1999)

Not freely available datasets

• UNB ICX IDS 2012 (Received authorization to use the dataset on August 1st therefore it will be used in a future version of the project). As I do not own the rights to it I cannot distribute it.

For the first part of the prototyping process, the KDD Cup dataset will be used as it was preprocessed already. It uses the following inputs and the goal is to categorize the example in one of the five categories.

Inputs

For now the project will be based on KDD's data which includes 40 values, 32 continuous, 8 of which are symbolic and will have to binarized if they aren't already. While some value are described as continuous in the dataset, it is important to note that it isn't the mathematical definition in the sense that { 0, 1, 2, 3, 4 } will be described as continuous.

When symbolic variables are binarized, the resulting vector contains 3 (protocol) + 70 (service) + 11 (flag) + 37 (continuous or already binarized) = 121 values

While all of those values will be used in the first version to allow comparison with current research, some might be added afterward for sake of completeness and relevancy in 2017.

Outputs

There are multiple possible outputs that can be grouped in

Considerations over the KDD Cup 99 dataset

While one of the few datasets freely available on the subject, the KDD Cup 99 dataset has many shortcomings that can be summarized to:

• Categories are unbalanced, 70% of the testing set is DoS
• Detecting DoS is overall pointless as it generates a log of traffic
• Redundant records

All those points and many more are described in A detailed analysis of the KDD Cup 99 Data Set by Mahbod Tavallaee, Ebrahim Bagheri, Wei Lu, and Ali A. Ghorbani. They proposed a solution which will be partially used namely:

• Removing redundant records
• Artificially increase the number of R2L and U2R examples

Model developpement

While researchers already explored most machine model, they forgot one aspect that is crucial to the project at hand: resources. The whole point being to end up with a usable hardware implementation, it is impossible to afford more than a few milliseconds for each packets on a rather weak machine. I will refrain from going straight to the LSTM and take the time to benchmark the performance of other well-known algorithms. I will explore three models to identify the one that shows the best performance to resources among logistic regression, deep feedforward neural networks, and LSTM recurrent networks.

To avoid reinventing the basics, I will use the TensorFlow framework that supports both x86-64 and ARM.

Logistic regression

• Pros
  • Fast inference
  • Simple to understand because it is not a blackbox
• Cons
  • Probably too simple to correctly fit the problem
  • No conception of time

Standard FeedForward Neural Network

• Pros
  • Complex enough to fit most problem
• Cons
  • No conception of time

LSTM Recurrent Neural Network

• Pros
  • Works with sequence, in our case, the packet that came before could be related to the next one. Such relation wouldn't be handled by the above models.
  • Documented use in IDS
• Cons
  • More complex than all three of the above.

Hardware implementation

Training

I will use my personnal computer for the training, if someone wishes to replicate the process, AWS offers VPS with GPUs .

• AMD Zen 1700
• Nvidia GTX 1080Ti

Inference

These are merely supposition on what minimal hardware would be required based on the platform. In the case at hand I will use an x86-64 processor because the cost for a dual gigabit port is lower and the hardware will be supported out of the box (I'd rather avoid troubleshooting hardware issues).

• x86-64

  PCPartPicker part list / Price breakdown by merchant

  Type	Item	Price
  CPU	Intel - Celeron G3930 2.9GHz Dual-Core Processor	$38.69 @ OutletPC
  Motherboard	Gigabyte - GA-H110M-A Micro ATX LGA1151 Motherboard	$51.49 @ SuperBiiz
  Memory	Patriot - Signature Line 4GB (1 x 4GB) DDR4-2400 Memory	$32.99 @ Amazon
  Storage	Seagate - Barracuda 250GB 3.5" 7200RPM Internal Hard Drive	$16.34 @ Other World Computing
  Case	Rosewill - SRM-01 MicroATX Mini Tower Case	$29.99 @ Amazon
  Power Supply	EVGA - 400W ATX Power Supply	$29.99 @ Amazon
  Wired Network Adapter	TP-Link - TG-3468 PCI-Express x1 10/100/1000 Mbps Network Adapter	$12.14 @ OutletPC
  Prices include shipping, taxes, rebates, and discounts
  Total	$211.63

  It is worth noting that a used server can probably do the job and cost less. Also I'd like to point out that this is a dirty cheap version, the parts can easily be changed to get a stronger machine, Celeron -> i3 or HDD -> SDD

• ARM

  • 4 cores ARMv8 or more recent
  • Something like a Banana Pi R2 hardware wise.

For portability and performance, the inference will be done in C++.

References

All consulted papers & documentation is available in the "papers" folder.