Replies to This Discussion

Permalink Reply by Dane on July 3, 2016 at 7:38pm

It would pass all autopens and fail most authentic signatures.

Permalink Reply by Paul on July 3, 2016 at 8:40pm

They use different machines for different lighting that reveals autographs that have been tampered with. So there are machines.

Permalink Reply by Paul on July 3, 2016 at 9:04pm

I think psa has this software on their quicj opinion service. You have to laugh.

Permalink Reply by Mike T on July 4, 2016 at 8:14am

It's definitely possible, from a tech standpoint. I've often thought about writing something like this, the computer could tell you how close a sample is to a known authentic exemplar, and if you had enough authentic exemplars it could, in theory, be pretty accurate. It would take quite a bit of work though, and it would never be able to distinguish live ink from preprinted signatures from just a scan.

Permalink Reply by Joe W. on July 4, 2016 at 9:17am

I believe it could be done. Would it be 100% accurate? Nope! Too many variables. I think nothing beats hands on experience. In the end, the only way anyone knows with absolute certainty is to get the signature in-person. 

Permalink Reply by Robin Moses on July 4, 2016 at 10:50am

Facial recognition is fairly simple.. The algorithm first looks for ANY face. Faces have very distinct shapes, and are easy to find. Then it measures fixed distances between facial features, and checks for matches.

For credit card transactions, your  signature is recorded with respect to time. It's very easy to verify temporal positioning, because the signature only exists at a single 2d point at any given time. "Offline signature verification" is much more difficult, because the time dimension gets compressed into a single point, leaving the signature as a projection. Now imagine trying to examine a signature that's been squeezed into a 1px line.

With machine learning, as is often the case, the research I've seen is questionable. The paper you attached claims 96% accuracy, but only reports their FINAL rate, after multiple rounds of "cheating" (section 5.2). There is no way their model would be as accurate with new data. They also used a data set with extremely obvious "first try" forgeries. I've attached examples. Writing a normal algorithm to find them would be trivial.

To be fair, they were only working with the data available to them, and were just doing research. With better data and more time,  it would be possible to develop actual useful models.

There are many things a model will detect better than a human. However, a model should only be used to inform a human, that way the human can make a better decision.

Attachments: No photo uploads here

• original_1_8.png, 100 KB
• forgeries_1_8.png, 32 KB
• forgeries_1_11.png, 31 KB

Permalink Reply by Virgil Thompson on July 4, 2016 at 11:26am

I agree. Use of every reliable tool available would help reduce the subjectivity and error

and fraud inherent to authentication. I doubt the first wheel was perfectly round.

 Hopefully the research continues.

Permalink Reply by Steve Zarelli on July 4, 2016 at 7:40pm

The problem is that the software would only be analyzing "shape," and most skilled fakes mimic the shape well within the range of natural variance. Good fakes are detected by atypical speed, flow and pressure... And sometimes they simply look "off" in a subtle, yet concerning way. And how would the software detect a trace job?

I believe software of this sort could be effective for detecting significantly malformed examples, but could it consistently flag skilled fakes? I doubt it.

Permalink Reply by Eric Keith Longo on July 4, 2016 at 8:11pm

What he said. And things like placement, and other less tangible things, many if not all of which are variable, the "flavor"...which ink for whom when as well...say it could be done, should it?

Eric

Permalink Reply by Robin Moses on July 4, 2016 at 10:42pm

A relatively simple algorithm will look at more that just the "shape" of an autograph.

I can identify forgeries in this data set (242 MB) using an algorithm that doesn't consider "shape" at all. It's actually easier that way.

I'm not sure what you mean by "atypical speed, flow, and pressure," but I believe that is what my algorithm is actually taking into account. Without going into too much detail, I'm comparing local intensity histograms at specific (calculated) points.