International Journal of Trend in Scientific Research and Development (IJTSRD)

Volume 8 Issue 5, Sep-Oct 2024 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470

Detection of AI-Generated Images

Misal Thakre1, Satyam Kadu2, Ronit Bera3, Rohit Atre4, Prof. Shreya Bhanse5
1,2,3,4
School of Science, G H Raisoni University, Amravati, Maharashtra, India
5
Assistant Professor, G H Raisoni University, Amravati, Maharashtra, India

ABSTRACT How to cite this paper: Misal Thakre |

Generative AI has gained enormous interest nowadays due to new Satyam Kadu | Ronit Bera | Rohit Atre |
applications like ChatGPT, DALL E, Stable Diffusion, and Prof. Shreya Bhanse "Detection of AI-
Deepfake. In particular, DALL E, Stable Diffusion, and others Generated Images" Published in
International
(Adobe Firefly, ImagineArt, etc.) can create images from a text
Journal of Trend in
prompt and are even able to create photorealistic images. Due to this Scientific Research
fact, intense research has been performed to create new image and Development
forensics applications able to distinguish between real captured (ijtsrd), ISSN:
images and videos and artificial ones. Detecting forgeries made with 2456-6470,
Deepfake is one of the most researched issues. This paper is about Volume-8 | Issue-5, IJTSRD69453
another kind of forgery detection. The purpose of this research is to October 2024,
detect photorealistic AI-created images versus real photos coming pp.805-810, URL:
from a physical camera. Id est, making a binary decision over an www.ijtsrd.com/papers/ijtsrd69453.pdf
image, asking whether it is artificially or naturally created. Artificial
images do not need to try to represent any real object, person, or Copyright © 2024 by author (s) and
International Journal of Trend in
place. For this purpose, techniques that perform a pixel-level feature Scientific Research and Development
extraction are used. The first one is Photo Response Non-Uniformity Journal. This is an
(PRNU). PRNU is a special noise due to imperfections on the camera Open Access article
sensor that is used for source camera identification. The underlying distributed under the
idea is that AI images will have a different PRNU pattern. The terms of the Creative Commons
second one is error level analysis (ELA). This is another type of Attribution License (CC BY 4.0)
feature extraction traditionally used for detecting image editing. ELA (http://creativecommons.org/licenses/by/4.0)
is being used nowadays by photographers for the manual detection of
AI-created images. Both kinds of features are used to train KEYWORDS: Computer Vision,
convolutional neural networks to differentiate between AI images Object Detection, Image
and real photographs. Good results are obtained, achieving accuracy Classification
rates of over 95%. Both extraction methods are carefully assessed by
computing precision/recall and F1 -score measurements. The
proliferation of AI-generated images, often referred to as deepfakes
or synthetic media, has revolutionized how digital content is created,
consumed, and shared. While these advancements offer immense
creative potential, they also pose significant challenges, particularly
in terms of authenticity and misinformation. This paper explores the
growing need for AI-generated image detection, especially in social
media applications, and delves into the methods used for
distinguishing between human-made and AI-generated content. It
outlines the technical challenges, the potential societal impact, and
strategies for implementing robust detection mechanisms in social
media platforms.

I. INTRODUCTION
Nowadays, generative artificial intelligence is one of authorship. Transformers were first used for
the top themes of computer engineering research. The automated translation systems, where a first
emergence of transformers [1] as a key tool for processing stage (the encoder) transforms the input
generating content has opened a world of new text into a numerical representation of text meaning;
applications where automated systems can create then a second stage (the decoder) converts (like an
productions that, until now, were exclusive to human inverse transform) those intermediate data into text in

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 805
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
another language [2]. Besides neural machine were found the most appropriate for photorealistic
translators, other impressive applications have arisen. images; other models are good at producing drawings
Famously, ChatGPT is a conversational engine or illustrations and not so much at imitating real
created with a decoder transformer [3]. Using photographs). The prompt was the same for the three
transformers, models for translating regular text into images: “realistic photo, a portrait of a dog in a
images have also been developed. The most known library, Sigma 85 mm f/1.4”. Note that details about
and documented examples of these last ones are the lens were added (85 mm focal lens, f/1.4 numeric
DALL E [4,5] and Stable Diffusion [6]. But other aperture); this is a common trick used for getting
examples have quickly been released, like OpenArt more realistic results. In the same figure, we also
[7], ImagineArt Adobe Firefly and many others. present three real photographs that will be processed
These artificial image generators have reached the later. The purpose of this work is to make a binary
point where they may create photorealistic images decision between two options: AI image (fully
that can make humans hesitate on whether a particular created AI image) and real image. Sensors 2023, 23, x
image is coming from a camera or is an artificial FOR PEER REVIEW 2 of 15 photorealistic images
creation. As an example, in Figure 1, three AI-created that can make humans hesitate on whether a particular
images are presented. They were created by three image is coming from a camera or is an artificial
different engines: DALL E 2, Stable Diffusion, and creation.
OpenArt (after testing many applications, these three

Another impressive AI application is Deepfake . Deepfake can create photos and videos mixing plausible
information from previous photos and/or videos. For example, creating a video of a person mixing the body of
one given individual and the face of another one. The potential danger of this technology being used for fraud or
other illegal purposes (defamation, pornography, etc.) has sparked much research in the field of detecting
Deepfake image creation . For example, in, Rössler et al. start by creating a large dataset of fake videos. In
authors exploit what is, perhaps, the most intuitive method: finding image artifacts that can reveal synthetic
content. In, a system called “FakeCatcher” is described; this system generated) images, which is another type of
problem as they deal with images that were created with intensive human intervention. For a similar need,
Google has recently announced a new tool called SynthID which adds an invisible watermark to AI-generated
images so that they can be identified. Note that this will identify AI images only if the creation engine
watermarks them. Because of the need to classify whole images with no assumption about image content, the
system was designed based on methods from other image forensics applications.
The main idea is to extract some relevant t information from images before applying a convolutional neural
network. Convolutional neural networks (CNNs) are very useful in distinguishing between classes that are
visually different for humans, like digit classification distinguishing objects relevant for making driving
decisions in real traffic, and many other similar applications Nevertheless, in this case, classes are not visually
different, and that suggests that the direct application of CNNs could not be very useful (besides the experience
from the Deepfake case). For this reason, pixel-wise feature extraction was used. This means using processing

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 806
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
stages that convert images into other images with the same size (it converts each pixel to a new pixel) but
containing a reduced amount of information that should be relevant to the particular problem of distinguishing
AI images.

Fig 2. Classification of Images

II. BACKGROUND
A. AI-Generated Image Techniques
The advancement of AI-generated images can largely be attributed to the development of Generative Adversarial
Networks (GANs). GANs consist of two neural networks: a generator that creates synthetic data and a
discriminator that evaluates whether the data is real or AI-generated. Over time, the generator improves its
ability to create realistic images, resulting in AI-generated visuals that are often indistinguishable from real
photographs.
In addition to GANs, models like DALL-E and Stable Diffusion have further refined image generation, using
text inputs to create complex visuals. These models have been used for everything from creating art to
generating images for advertisements, sparking both enthusiasm and concerns about their potential misuse.
B. The Rise of Deepfakes
Deepfakes are AI-generated media, often created using GANs, that replace the likeness of a person in a video or
image with someone else. They can be used for comedic or creative purposes but have also been weaponized to
create non-consensual imagery, fake news, or other forms of misinformation. The rise of deepfakes has triggered
the need for more sophisticated detection methods to prevent malicious use.
C. Challenges in Detection
Detecting AI-generated images is increasingly difficult as AI models become more advanced. While early AI-
generated images often contained visible flaws, modern systems can create visuals with highly realistic details
such as lighting, texture, and facial expressions. This poses a significant challenge for detection algorithms,
which must keep pace with rapidly evolving AI techniques. Moreover, AI-generated content can be modified to
evade detection, making it necessary for detection systems to be continuously updated.

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 807
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470

Fig 3. Challenges in detection

III. TECHNIQUES FOR AI-GENERATED C. Frequency Domain Analysis
IMAGE DETECTION The frequency domain refers to the representation of
A. Deep Learning-Based Detection an image in terms of its frequency components,
Deep learning models trained to recognize patterns typically obtained through Fourier or wavelet
unique to AI-generated images have emerged as a transforms. AI-generated images often contain
leading solution. These models analyze pixel-level frequency artifacts or irregularities that are not
anomalies or inconsistencies in texture, lighting, or present in naturally occurring images. By analyzing
edge detection that are more common in synthetic the frequency domain, detection systems can identify
images than in real ones. Convolutional Neural patterns that distinguish synthetic images from real
Networks (CNNs) are commonly used for this ones.
purpose due to their ability to extract complex This technique is particularly useful for detecting
features from images. fine-grained differences in textures and shading, but it
CNN-based detectors are trained on large datasets can struggle with more complex images or images
containing both real and AI-generated images, that have been post-processed to remove artifacts.
allowing them to learn the subtle differences between D. Blockchain and Digital Watermarking
the two. However, as AI-generated images become
One proposed solution to the problem of AI-
more sophisticated, deep learning models must be generated image detection is the use of blockchain
frequently retrained on new datasets to maintain technology or digital watermarking to verify the
accuracy. authenticity of images. Blockchain could be used to
B. Analysis track the provenance of an image from its original
Another method for detecting AI-generated images creation, ensuring that any modifications or AI
involves analyzing the metadata associated with the alterations are recorded. Digital watermarking
image file. AI-generated images often lack specific involves embedding invisible information into an
metadata or contain metadata that is inconsistent with image that can be detected to verify its authenticity.
that of genuine photos, such as information about the While promising, these techniques require widespread
camera or software used. Detection algorithms can adoption and cooperation from content creators and
flag these inconsistencies as potential indicators of platforms to be effective. Additionally, they do not
synthetic content. address the detection of AI-generated images that
However, this approach has limitations, as metadata have not been watermarked or tracked via blockchain.
can be easily altered or removed, making it less
reliable as a standalone detection method.

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 808
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
IV. IMPLEMENTATION IN SOCIAL V. CONCLUSION
MEDIA PLATFORMS The rise of AI-generated images presents both
A. User Experience And Integration opportunities and challenges for social media
The successful implementation of AI-generated platforms. While these technologies enable creative
image detection in social media applications requires expression and innovation, they also pose significant
seamless integration into the platform's existing risks in terms of misinformation and trust.
infrastructure. Users should be able to trust that the Implementing AI-generated image detection systems
platform is filtering out or flagging suspicious content is crucial for ensuring the authenticity of content on
without compromising their experience. Features such social media platforms. By using advanced detection
as real-time content analysis, automated flagging, and techniques such as deep learning, frequency analysis,
transparency reports on detected content should be and metadata examination, platforms can better
made available to users and moderators alike. protect users from deceptive content and maintain the
Social media platforms can also give users tools to integrity of their ecosystems.
verify images themselves, such as "Verify Image" As this field continues to evolve, ongoing research
buttons or notifications that inform users when a and collaboration between developers, regulators, and
piece of content has been flagged as potentially AI- platform operators will be essential to creating a safer
generated. and more trustworthy online environment.
B. Ethical Considerations This paper could serve as a strong foundation for your
While AI-generated image detection is necessary to research or product development. Let me know if
prevent misinformation, there are ethical concerns you'd like to focus on any specific areas for further
that must be addressed. Algorithms must be exploration or discuss potential implementation
transparent in how they detect AI-generated content strategies for your application.
to avoid bias or censorship, and users should be
VI. REFERENCES
informed about why certain images are flagged.
[1] Usha Kosarkar, Gopal Sakarkar, Shilpa Gedam
Additionally, false positives could lead to real, user-
(2022), “An Analytical Perspective on Various
generated content being unfairly flagged or removed,
Deep Learning Techniques for Deepfake
necessitating a balanced approach that includes
Detection”, 1st International Conference on
human oversight.
Artificial Intelligence and Big Data Analytics
C. Legal and Regulatory Framework (ICAIBDA), 10th & 11th June 2022, 2456-3463,
The rise of AI-generated images has prompted Volume 7, PP. 25-30
discussions around the need for new regulations to
[2] Usha Kosarkar, Gopal Sakarkar, Shilpa Gedam
govern their use, particularly in social media.
(2022), “Revealing and Classification of
Governments are exploring the idea of mandatory
Deepfakes Videos Images using a Customize
disclosure for AI-generated content, which would
Convolution Neural Network Model”,
require platforms to label synthetic media. By
International Conference on Machine Learning
integrating detection systems, social media companies
and Data Engineering (ICMLDE), 7th & 8th
can stay ahead of regulatory changes while promoting
September 2022, 2636-2652, Volume 218, PP.
responsible content sharing.
2636-2652,
D. Future Directions https://doi.org/10.1016/j.procs.2023.01.237
As AI-generated images become more difficult to
[3] Usha Kosarkar, Gopal Sakarkar (2023),
detect, new approaches will be needed to keep
“Unmasking Deep Fakes: Advancements,
detection systems effective. Hybrid methods that
Challenges, and Ethical Considerations”, 4th
combine deep learning, frequency analysis, and
International Conference on Electrical and
metadata analysis may offer the most robust
Electronics Engineering (ICEEE),19th & 20th
solutions. Research into adversarial machine learning, August 2023, 978-981-99-8661-3, Volume
where detection systems are trained to combat AI 1115, PP. 249-262, https://doi.org/10.1007/978-
systems designed to evade them, will also be critical 981-99-8661-3_19
to staying ahead of emerging threats.
[4] Usha Kosarkar, Gopal Sakarkar, Shilpa Gedam
Moreover, the development of decentralized (2021), “Deepfakes, a threat to society”,
verification systems, such as blockchain-based International Journal of Scientific Research in
content tracking, could offer a long-term solution to Science and Technology (IJSRST), 13th October
the challenges posed by AI-generated media. 2021, 2395-602X, Volume 9, Issue 6, PP.
1132-1140, https://ijsrst.com/IJSRST219682

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 809
 International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
[5] Usha Kosarkar, Prachi Sasankar(2021), “ A Applications, 1380-7501,
study for Face Recognition using techniques https://doi.org/10.1007/s11042-024-19220-w
PCA and KNN”, Journal of Computer
[7] Usha Kosarkar, Dipali Bhende, “Employing
Engineering (IOSR-JCE), 2278-0661,PP 2-5, Artificial Intelligence Techniques in Mental
[6] Usha Kosarkar, Gopal Sakarkar (2024), Health Diagnostic Expert System”,
“Design an efficient VARMA LSTM GRU International Journal of Computer Engineering
model for identification of deep-fake images (IOSR-JCE),2278-0661, PP-40-45,
via dynamic window-based spatio-temporal https://www.iosrjournals.org/iosrjce/papers/con
analysis”, Journal of Multimedia Tools and f.15013/Volume%202/9.%204045.pdf?id=7557

@ IJTSRD | Unique Paper ID – IJTSRD69453 | Volume – 8 | Issue – 5 | Sep-Oct 2024 Page 810