Thread by @eh_den

IS #AI ABOUT TO STEAL THE 2024 ELECTION?

1/[THREAD]

In order to steal the upcoming election, all one needs to do is to find a way to perfectly fake the handwritten signatures of potential voters.
Sadly, it seems extremely likely this is possible, with the use of AI.
#StopAIVote

2.
One of the biggest problems of mail-in voting is signature verification. To resolve that, there is a requirement to perform a signature verification.

CISA's "Signature Verification and Cure Process.", details how you verify signatures.

#StopAIVote
cisa.gov/sites/default/…

3.
Mail ballot sorting equipment and voter registration system use Automated signature verification (ASV) applications that use a camera to capture the voter’s signature from the ballot return envelope & compared with reference image from voter registration database.
#StopAIVote

4.
According to the CISA guide, the usage of signature verification software is considered "Tier 1".

It means that if the software provided a "correct" answer, the ballot is not required to pass any further verification.

Fool the AVS and you own the election.

#StopAIVote

5.
So, what's AI has to do with all of this?

Let's assume you want to rig the election using main-in ballots of either fake voters or even real voters, and make sure that no one will be able to catch you, NOT EVEN VIA A HUMAN INSPECTION.
You need AI for that.
#StopAIVote

6.
HOW TO USE AI TO RIG A VOTING SYSTEM THAT ALLOWS MAIL-IN BALLOTS.
1) Obtain a signature verification database.
2) Train AI/ML on the signatures.
3) Train an AI robot printing system to print a replication of signatures.
4) You own the election.
#StopAIVote

7.
What I described in #6 is far from hypothetical.

EXAMPLE 1:
Bond, a 2015 startup in New York City created robots that can learn your handwriting and convert digital notes into personalized letters. It also worked with signatures.

nytimes.com/2015/12/17/bus…

#StopAIVote

8.
EXAMPLE 2:
The Handwriting Company: a start-up that aimed of using ML/AI to mimic handwriting by scanning the handwriting. "It's not just printing the words on a paper. It's applying pressure at certain points I apply pressure".
#StopAIVote

9.
OK, so there might be some systems out there that were able to replicate handwriting.
But signatures are a bit different, and Bond which allowed signature replication had a calligraphers who provided the quality control.
Let's first look at what is a signature.
#StopAIVote

10.
In "Modeling the Lexical Morphology of Western Handwritten Signatures", Diaz-Cabrera et al focused on the lexical morphology of handwritten signatures, meaning the identification, analysis, and description of the signature structures.
#StopAIVote
doi.org/10.1371%2Fjour…

11.
According to the paper, Western signatures are generally composed of text (reflects the name, the family name etc) plus a flourish, or rubric in the occidental signatures which is defined by a kind of doodle written much faster and without much attention.
#StopAIVote

12.
The researchers used a small set of possible parameters to gain a better understanding of the main factors which characterize the way the signatures are performed. But here are multiple parameters which construct a signature.

Let's look at verification methods.
#StopAIVote

13.
In "Deep Learning Based Handwritten Signature Recognition" (Pokharel et al, 2020), the authors used deep learning model based on the CNN architecture to verify the signature with 95.2% accuracy.
#StopAIVote

researchgate.net/publication/35…

14.
In “Offline signature identification via HOG features and artificial neural networks" (M. Taxkiran and Z. G. Çam, 2017), the authors demonstrated 98.33% accuracy rate.
#StopAIVote
doi.org/10.1109/SAMI.2…

15.
In "Offline Signature Verification Using Convolutional Neural Network" (Bonde et al, 2022) the authors used Gaussian Weighting Based Tangent Angle (GWBTA)+fine tuned convolutional neural network (CNN) and used it to classify authenticity.
#StopAIVote
doi.org/10.1109/ICSC56…

16.
I gave the examples above to show there are many ways to validate signatures, with increased accuracy, which can be deployed in Automated signature verification (ASV) systems.
Which brings us back the the question: can a machine learn to perfectly fake signatures?
#StopAIVote

17.
In "Writer-independent signature verification; Evaluation of robotic and generative adversarial attacks" (bird et at, 2023) authors demonstrated that using a XY robotic pen plotter (iDraw 2.0 and line-us) it is possible to identify forgery.
#StopAIVote
doi.org/10.1016/j.ins.…

18.
What about XYZ plotters?
In 2019, "Design of XYZ Plotter" (Kumar et al, 2019), the authors suggested a low-cost X-Y-Z plotter using Arduino as the controller.
engineeringjournals.stmjournals.in/index.php/JoCI…

19.
In 2019 a Brown University undergraduate student and a staff member developed an algorithm that enables robots to reproduce human-like pen strokes just by looking at images of handwriting or sketches, with 93 percent accuracy,

#StopAIVote
brown.edu/news/2019-05-1…

20.
In the paper, entitled "Teaching Robots To Draw", the authors shown that their system was able to replicate characters from multiple languages which it was not trained upon.

h2r.cs.brown.edu/wp-content/upl…

#StopAIVote

21.
They also stated that "In the future, we will explore more advanced types of drawing such as drawing with a paint brush, where the speed of the stroke, the height/altitude/azimuth of the brush relative to the canvas plane all affect the drawing result." >>
#StopAIVote

22.
>> "In this vein we are interested in closing the learning loop by enabling the robot to inspect what it has drawn and improve its drawing policy based on observation of what it actually drew".

#StopAIVote

23.
Based on the above, I suggest that the following model will result in an impossible to identify forgery.
1) Get a large database of offline and online signatures databases. See
dx.doi.org/10.1145/3274658
2) Train AI on databases using multiple approaches.
#StopAIVote

24.
3) Use a robotic XYZ print plotter with azimuth, pressure and speed capability to try and replicate signatures.
4) Capture the signature as a new offline signature database.
5) Use multiple signature verification algorithms to score the accuracy.

#StopAIVote

25.
6) Based on the score, to re-train the system, input to the offline database and train it until it reaches the required accuracy.

The above scenario is to plausible and feasable to be ignored. IMHO, Due to the progress in AI, mail-in ballots should be prohibited.
#StopAIVote