IDENTIFICATION OF IMAGES
GENERATED BY ARTIFICIAL INTELLIGENCE
“TERESIUS-AI”
User Manual
2024
Table of Contents
- INTRODUCTION
1.
DESCRIPTION OF THE
PHYSICO-MATHEMATICAL PRINCIPLES OF IDENTIFICATION
2.
WORKING WITH THE SYSTEM
3.
GRAPHS OF TYPE I AND TYPE II ERRORS
INTRODUCTION
This version (TERESIUS_AI.exe) of the software complex “TERESIUS-AI”
(hereinafter referred to as the System) is designed to
solve the problem of IDENTIFYING ARTIFICIAL INTELLIGENCE (AI) BASED ON VIDEO
IMAGES GENERATED BY IT (Web version of the system for individual users).
Currently, there are several systems that address
similar expert assessment tasks. The proposed version of the system, based on
preliminary testing data, demonstrates high effectiveness in terms of
decision-making error rates (0.01-5%). The
software system aims to determine whether a given image was
created by an artificial intelligence application or generated without AI
assistance. Additionally, it provides a probability of
decision-making error, which is a variable that depends on the specific AI
application, the nature of the image, and various other factors.
The system operates for any resolution of video
matrices exceeding 100x100 pixels, specifically for color video recordings in
RGB format of any kind. It is not necessary to select video frames with
identical resolutions for identification. The system will analyze any video
files and frames of varying resolutions (greater than 100x100) in any graphical
format.
1. DESCRIPTION OF THE
PHYSICO-MATHEMATICAL PRINCIPLES OF IDENTIFICATION
The
ideology of the identification system at the physical level is as follows. Any
image generated by artificial intelligence is created by a specific software
application (such as DALL-E, Bin, Midjourney, etc.)
based on a defined sequence of algorithms. These are different algorithms,
although their construction principles may be common. An image generated by
these algorithms exhibits certain statistical patterns of color palette
transitions between pixels and possibly numerous other individual
characteristics specific to that application.
These characteristics are, in
most cases, not visually distinguishable by humans. At the same time,
regardless of the image generation algorithms employed by AI, these algorithms
cannot be identical to those by which nature forms our surrounding world (at
least at the current level of scientific understanding).
Let us
denote the entire set of integral characteristics of the algorithms of a
specific AI application as the “Artistic Style of the AI Application.” In this
context, this style bears no relation to the conventional understanding of
style, such as that of artists, renowned photographers, or any human
expressions of style.
The
characteristics of such AI-generated image styles are virtually impossible to
perceive visually. To
highlight the statistical individual features of the AI application style, it
is necessary to employ a mathematical framework for image processing that
allows for the quantification of style characteristics. In this system, several
combinations of wavelet decomposition of images are utilized
as such a mathematical apparatus. Our long-term research has shown that several
parameters of these mathematical transformations are
correlated with the algorithms of AI applications that generate images.
The verbal exposition of the mathematical approach is, of course, a hypothesis.
The effectiveness of this hypothesis can only be confirmed
by the performance of the developed identification system for AI applications.
After extracting parameters from
the wavelet decomposition of images, a specific function of these parameters
for a given image is formed in the system. This
function served as input during the training of a deep learning neural network
for binary image identification. The identification criterion is whether the
image was CREATED BY ARTIFICIAL INTELLIGENCE or NOT. The neural network was trained on several
hundred thousand different images. The process of augmenting the dataset for
further training continues (evolving the neural network model and its
modification). The resulting neural network model is applied
to identify images generated by artificial intelligence.
2. WORKING WITH THE SYSTEM
The operation of the system is initiated on the website through the TERESIUS_AI option. The user will be presented
with a file selection option for analysis (see Figure 1).
Fig 1. File Selection Window for
Analysis
After selecting a file, the user
must choose the load option. Upon completion of the analysis, a graphical
window will open displaying the analysis results (see Figure 2).
Fig 2. Identification Results
The error probability presented
after calculations on the graph is a variable quantity. This probability
depends on the AI application that generated the image, the resolution and
nature of the image, and various other factors.
When reanalyzing the same image,
this probability value may differ slightly. Such a difference results from the
image analysis technology. In this analysis, templates of images created by AI
for specific applications are occasionally used, and these templates are selected randomly each time. However, this technology
does not significantly affect classification results.
3. GRAPHS OF TYPE I AND TYPE II ERRORS
The graphs of Type I and Type II
errors for the current efficiency of the system are presented
in Figure 3. These graphs were obtained through
testing the system on a large volume of statistical data.
Fig 3. Graphs of Type I and Type
II Errors
The x-axis represents the value
P, which is the calculated probability of error for a specific image based on
the neural network model. The values for Type I and Type II errors on the
graphs are derived from test trials of the system on
independent test data for a large volume of photographic images.