Adapting Machine Learning Algorithms to my Problem

i'm working on a project and need your ideas, advices.

First of all, let me tell my problem.

There is power button and some other keys of a machine and there is only one user has authentication to use this machine.There are no other authentication methods, the machine is in public area in a company.

the machine is working with the combination of pressing power button and some other keys. The order of pressing keys is secret but we don't trust it, anybody can learn the password and can access the machine.

i have the capability of managing the key hold time and also some other metrics to measure the time differences between the key such as horizantal or vertical key press times (differences). and also i can measure 开发者_Python百科the hold time etc.

These all means i have some inputs,

Now i'm trying to get a user profile by analysing these inputs.

My idea is to get the authenticated user to press the password n times and create a threshold or something similar to that.

This method also can be said BIOMETRICS, anyone else who knows the machine button combination, can try the password but if he is out of this range can not get access it.

How can i adapt these into my algorithms? where should i start ?

i don't want to delve deep into machine learning, and also i can see that in my first try i can get false positive and false negative values really high, but i can manage it by changing my inputs.

thanks.

To me this seems like a good candidate for a classification problem. You have two classes (correct password input / incorrect), your data could be the time (from time 0) that buttons were pressed. You could teach a learning algorithm but having several examples of correct password data and incorrect password data. Once your classifier is trained and working satisfactorily, you could try it out to predict new password input attempts for correctness.

You could try out several classifiers from Weka, a GUI based machine learning tool http://www.cs.waikato.ac.nz/ml/weka/

What you need is your data to be in a simple table format for experimenting in weka, something like the following:

Attempt No | 1st button time | 2nd button time | 3rd button time | is_correct
-----------|-----------------|-----------------|-----------------|------------
     1     |    1.2          |    1.5          |  2.4            | YES
     2     |    1.3          |    1.8          |  2.2            | YES
     3     |    1.1          |    1.9          |  2.0            | YES
     4     |    0.8          |    2.1          |  2.9            | NO
     5     |    1.2          |    1.9          |  2.2            | YES
     6     |    1.1          |    1.8          |  2.1            | NO

This would be a training set. The outcome (which is known) is the class is_correct. You would run this data through weka selecting a classifier (Naive Bayes' for example). This would produce a claffier ( for example a set of rules) which could be used to predict future entries.

The key to this sort of problem is devising good metrics. Once you have a vector of input values, you can use one of a number of machine learning algorithms to classify it as authorised or declined. So the first step should be to determine which metrics (of those you mention) will be the most useful and pick a small number of them (5-10). You can probably benefit by collapsing some by means of averaging (for example, the average length of any key press, rather than a separate value for every key). Then you will need to pick an algorithm. A good one for classifying vectors of real number is Support vector machines - at this point you should read up on it, particularly on what the "kernel" function is so you can choose one to use. Then you will need to gather a set of learning examples (vectors with a known result), train the algorithm with them, and test the trained svm on a fresh set of examples to see how it performs. If the performance is poor with a simple kernel (e.g. linear), you may choose to use a higher dimensional one. Good luck!