05.06.2013 Views

PNNL-13501 - Pacific Northwest National Laboratory

PNNL-13501 - Pacific Northwest National Laboratory

PNNL-13501 - Pacific Northwest National Laboratory

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Unknowns include<br />

1. What is a representative set of data?<br />

2. How do you know when training is complete?<br />

3. How do you know that the artificial neural networks<br />

is fully trained?<br />

4. How do you know that the artificial neural networks<br />

has not found local minima?<br />

5. How do you know that the artificial neural networks<br />

will generalize properly?<br />

Genetic Algorithms and Evolution Programming<br />

Genetic algorithms and evolution programming are<br />

unlikely to be useful in this complex near real-time<br />

reactive environment. However, genetic algorithms have<br />

been used successfully in optimizing neural networks.<br />

They may also be used to support new procedures<br />

required in a dynamic environment.<br />

Lattice Automata<br />

Lattice automata is a relatively new field being developed<br />

for the British defense community. Preliminary results<br />

show that a particular strength of this method is the ability<br />

of several autonomous agents to plan a simultaneous<br />

meeting time and location even in the event that the<br />

domain topology changes in a dramatic way. These<br />

agents will then be able to renegotiate a new plan. This<br />

strength could be used to coordinate several other<br />

decision-making engines, but not as part of the anomaly<br />

detection engine core.<br />

Artificial Intelligence Planning Systems<br />

Generative planning systems are computing intensive and<br />

are usually not considered in near real-time scenarios.<br />

Petri Nets and Colored Petri Nets<br />

Petri nets were developed to overcome the limitations of<br />

finite state machines, which could handle only a single<br />

event at any given time. Coordinated actions are difficult<br />

to model. Petri nets are good at describing actions, but<br />

are not good at describing the reasons behind the actions.<br />

Petri nets have been used to model network protocols and<br />

information flow. Other applications include<br />

152 FY 2000 <strong>Laboratory</strong> Directed Research and Development Annual Report<br />

1. Naval Command and Control System<br />

2. Intelligent network simulation<br />

3. Work flow model in a nuclear waste plant<br />

4. Model for document storage system<br />

5. Purdue IDIOT intrusion detection system.<br />

At the level of this research, petri nets are not applicable.<br />

Petri nets are not suitable for detection of novel and new<br />

threats (Bace 2000).<br />

Enhanced Anomaly Detection Engines<br />

With the multiple algorithm solution in mind, the<br />

enhanced engines will take advantage of any algorithm<br />

specially suited to a particular scenario that is presented.<br />

The decision engine allows from one to all available<br />

engines to be engaged upon the general engine detecting<br />

anomalous behavior. This allows us to implement<br />

methodologies that span the whole solution spectrum with<br />

minimum impact to the system. The following algorithms<br />

are considered top candidates for these engines.<br />

Operational model (Bace 2000) – event counters<br />

• Quantitative analysis (Bace 2000)<br />

• Threshold detection<br />

• Heuristic Threshold detection<br />

• Signature Analysis<br />

• Target Based Integrity Checks<br />

• Quantitative Analysis and Data Reduction<br />

Statistical Methods<br />

• Mean and Standard Deviation (Bace 2000)<br />

• Multivariate (Bace 2000)<br />

• Markov process (Bace 2000)<br />

• Parametric statistical methods<br />

• Non-parametric statistical methods<br />

Neural Networks<br />

• Neural networks will be used for one or more of the<br />

enhanced engines<br />

Rule Base Approaches<br />

• Time Based Induction Machine<br />

• Markov.<br />

Obviously, not all of these methods will be implemented.<br />

A subset of the above methods will be selected on the

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!