Numenta 2014 Spring NuPIC Hackathon Demos, 8 May 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Single-Entry Multifactor Authentication: 2014 Spring NuPIC Hackathon Demo

Description

Wade Hought

A

Unfortunately, I used the same computers, Marcos Marcus did so there's nothing really to hook up and show what I took was what I did was took the sign example and used it to recognize a password, but further than that recognize who is typing the password by measuring the latency between the keys and it basically seemed to work as long as the longer the passphrase got, the more some of the anomaly scores were harder to average out so I started.

A

Looking at your your method that you talked about and that started working, but there was some complexity in there, so I, that's kind of where I stopped, but anyway.

B

A

It was kind of cool like recognize who's typing the password rather than just the password. You.

B

Should mention the fields that you used? I used the.

A

Character just as a category and then I used the time latency, instead of actually using a never incremented time, I used the latency of time between each one of the characters.

A

So this would be you know, 70 milliseconds and 210 milliseconds and then obviously, if someone else typed it those time, gaps should be different and Marcus helped me a little bit and then he was kind of trying to type faster and see if we could catch it, but anyway and I might I might extend it and see if I could use like IP address MAC address time of day, maybe record how long the key is down and then how long? It's up you know heavy typers like typers, something like that. So.

B

I thought this was interesting just because it's kind of a fingerprint for how someone uses a keyboard to.

A

B

Extra security measure, in addition to the keys typed on a keyboard for typing in.

A

Your password well that's kind of where I came at it from us. My regular work has to do with finance stuff and multi-factor authentication is a complete pain in the butt. You know, because you've got a password, then you've got a key fob or you've got something else, and there's always something physical to to handle that if you could get to authentication factors in the same mechanism, that could be it.

B

Would be really hard to fake too, so you know when they do the retina scan and the movies, and then the bad guy takes the guys I out and just puts it up through I'd.

A

Be a lot harder, it would mean I. We want some people, you have it something anyway, but it was just off a basic sign example, which was kind of cool for me, because I didn't have new pic working until Saturday morning, one when we started it's. The sign example: I basically I mean I added some stuff to record when the keys went down and went up and recorded the data, but the model is exactly the same as the sign. I Reese warmed it a little bit.

A

I also took when I swarmed it the the spread of bits for the scalar for the time it mapped out and it was able to solve it and I think it came out with like a hundred and hundred and forty-four an and maybe 24 w I noticed that if I bump the an up after I'd sworn to like I, think six or seven hundred the resolution got way better in my anomalies course flattened out so like the swarmer was able to solve it at one level, but it you know once we actually went into detection mode, it couldn't kind of keep up with it.

A

I guess yeah probably I mean because I was doing it on like a six and eight character. Password then I tried some longer pass phrases and stuff, but yeah, ideally you'd want to train it on what that one person types over the space of days or weeks. So you could get you'd know exactly that. It's that person typing but I.

B

Get the impression that swarming isn't necessarily optimized for anomaly detection is that true I mean.

C

Yeah the waste warming works. Is it optimizes the prediction, error and.

A

C

Hoping that a model that is good at predicting will also be good at anomalies, but I think as we saw in some of the other examples. That's not always exactly true, but it's it's a first approximation, but what we would need to have swarming up for anomaly. Detection has come up with some metric, that's specific to anomaly detection and that that's kind of hard.

B

Any other questions for wait, alright, cool. Thank you.

B