Hi Rajesh,

The first thing you need to consider is what it is you want to compare between your test subjects. What question(s) are you trying to answer? You need to define that before you can work out how to compare the outputs.

Regards,

Cameron

Thank you, Dr. Wilson for responding to the question.

By comparing the signals generated by subject A and B, I want to find out whether the signals produced by subject A are close enough to that of subject B such that subject A can deceive an authentication system built for subject B. The constraint in my case is that I don't know the details of the authentication system and I am not allowed to send Subject A in place of Subject B unless I am sure that Subject A would successfully deceive authenticator build for Subject B. All I have is patterns stolen from subject B, and a skilled imitator subject A.

Hope that is the information you were looking for to answer my question.

That's an interesting question!

I'm not an expert, so others will have more comprehensive answers. But an approach I've used previously for accelerometer data is a Fast Fourier Transform analysis. This essentially gives you a plot of frequency (x axis) vs amplitude (y axis), which can identify dominant frequencies. I expect you could do this for each of the orthogonal directions. Although I'm not familiar with them, I expect there are plenty of statistical methods used in signal processing that would then allow you to compare the plots. It would be interesting to see if the dominant frequencies are similar; it may well be that the tell-tale differences will be in the less dominant frequencies.

It may also be interesting to monitor the signals from each direction relative to each other, but how best to do that is beyond my knowledge (I haven't done this sort of thing in almost 20 years).

One more thing to consider: the phone may be carried in a different place by each person, which might mean having to eliminate any strong signals from a swinging motion.

A source you might look to for more info would be the fitness tracker device industry. I imagine a lot of their processing algorithms are proprietry, but many may be available in patents or, if they have spun off from academic research, in related human movement literature.

Good luck!

Thanks again, for taking time to write the answer. You certainly provided some insight. Will follow the suggested direction and if I come up with something, I will make sure that I post that here. Thanks!

Here is the link of my article ,Hope is helpful :

Article SVM Multi-Classification of Induction Machine's bearings def...

Hi Rajesh,

I have done a simple experiment 2 years ago by calculating the Euclidean Distance between N number of users. As you know, if the distance between User A and User B is small that means there is a possibility that user A would be authenticated as user B and vice versa.

Another way is to calculate some time domain features such as Min, Max, STD and draw a figure ( but that one is a bit ugly as it does not show how these features are correlated to each other).

Kind Regards.

Hi Rajesh, I hope I am not too late to help, might I suggest cross-correlation (https://www.mathworks.com/help/signal/ref/xcorr.html). This seems like it may fit the bill.

A paper my colleagues and i published in the journal, Physiological measurement, may be of use too, https://www.researchgate.net/profile/Cain_Clark/publication/324038891/inline/jsViewer/5aba51d6aca2728f4fa38203 albeit this was done on magnetometer traces, the the analyses would correspond.

Best wishes, Cain