Beyond recognition

Posted on 19 Jan 2007 at 16:31

 

Gallery

One key consideration was how long a person had to walk around until the software could get a fix on his or her identity. In this test, recognition speeds were fast. By taking 256 slices of data per second from the accelerometer, the team found that the software could identify individuals after just 20 metres of walking. So if someone tried to walk off with your phone, by the time they were around the corner the phone would know it wasn't you carrying it and either shut down or call for help - at least until you used a less 'calm' means of identification, such as keying in a PIN, for instance.

M?yj?i sees uses for his discovery not only in smartphones but also in mobile banking and credit card payment applications. However, it will be some time before such protection becomes available, as there are still hurdles to overcome. "The potential drawbacks of the method are partly common to all gait-based methods," he warned. "Effect of changes in the speed of walking, [high heels] and ground; also drunkenness and injuries affect gait." M?yj?i also said he needed to understand the effect of carrying the device on different parts of the body. The data gathered from a phone swinging along in a bag might be very different from that gathered when it's in a coat pocket.

Internet use analysis

Other groups are investigating 'calm' methods of identifying us to the systems we use, such as websites. Detecting who's online from their unique usage patterns, for instance, might save the billions lost every year to online fraud, but only if such signatures are clear enough to give a positive identification first time every time.

One group of researchers believes these signatures do exist, and that they know how to spot them. Researchers Balaji Padmanabhan and Yinghui Yang have come up with clickprints, which they say are analogous to fingerprints. Clickprints, they claim, are capable of spotting someone online in a mass of visitors after just a few visits. In a complex online paper, Padmanabhan, of the Wharton School at the University of Pennsylvania, and Yang, of the University of California at Davis, revealed an algorithm that takes as its inputs things such as the time of access, source IP address, referring page, number of webpages visited and the time spent on each page. The output is a prediction of who is doing the surfing.

To determine its accuracy, Padmanabhan and Yang tested it using website usage data from comScore Networks. In much the same way that market research companies gather TV viewing habits, comScore provides real-time measurements of internet use, all captured from a consenting panel of one million users worldwide. For their tests, the researchers used data for the surfing habits from 50,000 comScore users. They also selected five popular sites, which remain anonymous in their results. From these 50,000 visitors, they chose just five for an initial test. The researchers selected five features from the site usage data. These were the duration of a visit to a site, the number of pages viewed, the average time spent on each page, the time the user visited each page and the day of the week on which the visit took place.

Surf's up

Using an open-source analysis program called Weka J4.8, the researchers wrote scripts to implement their algorithm. By analysing increasing numbers of surfing sessions for each user, they could gradually calculate the minimum number of surfing sessions needed to identify any of the five accurately enough to be of use. After just seven sessions, the algorithm could identify the users accurately nearly 87 per cent of the time. After 10 sessions, this rose to nearly 92 per cent. By 51 sessions, this was as high as 99.5 per cent. They also found that as the number of users of a site increased, the number of sessions needed to identify a user accurately rose, but not by much.