Received 15 December 2023, accepted 17 January 2024, date of publication 19 January 2024, date of current version 1 February 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3356122

CIFAKE: Image Classification and Explainable

Identification of AI-Generated Synthetic Images
JORDAN J. BIRD AND AHMAD LOTFI , (Senior Member, IEEE)
Department of Computer Science, Nottingham Trent University, NG1 4FQ Nottingham, U.K.
Corresponding author: Jordan J. Bird ([email protected])

ABSTRACT Recent advances in synthetic data have enabled the generation of images with such high
quality that human beings cannot distinguish the difference between real-life photographs and Artificial
Intelligence (AI) generated images. Given the critical necessity of data reliability and authentication, this
article proposes to enhance our ability to recognise AI-generated images through computer vision. Initially,
a synthetic dataset is generated that mirrors the ten classes of the already available CIFAR-10 dataset with
latent diffusion, providing a contrasting set of images for comparison to real photographs. The model is
capable of generating complex visual attributes, such as photorealistic reflections in water. The two sets of
data present as a binary classification problem with regard to whether the photograph is real or generated
by AI. This study then proposes the use of a Convolutional Neural Network (CNN) to classify the images
into two categories; Real or Fake. Following hyperparameter tuning and the training of 36 individual
network topologies, the optimal approach could correctly classify the images with 92.98% accuracy.
Finally, this study implements explainable AI via Gradient Class Activation Mapping to explore which
features within the images are useful for classification. Interpretation reveals interesting concepts within
the image, in particular, noting that the actual entity itself does not hold useful information for
classification; instead, the model focuses on small visual imperfections in the background of the images.
The complete dataset engineered for this study, referred to as the CIFAKE dataset, is made publicly
available to the research community for future work.

INDEX TERMS AI-generated images, generative AI, image classification, latent diffusion.

I. INTRODUCTION
and truth. This has led to a situation where consumer-level
The field of synthetic image generation by Artificial Intel-
technology is available that could quite easily be used for
ligence (AI) has developed rapidly in recent years, and the
the violation of privacy and to commit fraud. These
ability to detect AI-generated photos has also become a
philosophical and societal implications are at the forefront
critical necessity to ensure the authenticity of image data.
of the current state of the art, raising fundamental questions
Within recent memory, generative technology often
about the nature of trustworthiness and reality. Recent
produced images with major visual defects that were
technological advances have enabled the generation of
noticeable to the human eye, but now we are faced with the
images with such high quality that human beings cannot tell
possibility of AI models generating high-fidelity and
the difference between a real-life photograph and an image
photorealistic images in a matter of seconds. The AI-
that is no more than a hallucination of an artificial neural
generated images are now at the quality level needed to
network’s weights and biases.
compete with humans and win art competitions [1].
Generative imagery that is indistinguishable from pho-
Latent Diffusion Models (LDMs), a type of generative
tographic data raises questions both ontological, those
model, have emerged as a powerful tool to generate
which concern the nature of being, and epistemological,
synthetic imagery [2]. These recent developments have
surrounding the theories of methods, validity, and scope.
caused a paradigm shift in our understanding of creativity,
Ontologically, given that humans cannot tell the difference
authenticity,
between images from cameras and those generated by AI
models such as an Artificial Neural Network, in terms of
The associate editor coordinating the review of this manuscript and
digital information, what is real and what is not? The
approving it for publication was Yiqi Liu .
15642
©
2024
The
Author
s. This
work is
license
d
under
a
Creativ
e
Comm
ons
Attribu
tion
4.0
License
.
Fo
r
m
or
e
inf
or
m
ati
on
,
se
e
ht
tp
s:/
/cr
ea
tiv
ec
o
m
m
on
s.
or
g/l
ic
en
se
s/
by
/4.
0/
V
O
L
U
M
E
12
,
20
24
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

epistemological reality is that there are serious questions an individual committing a crime or vice versa,
surrounding the reliability of human knowledge and the providing false
ethical implications that surround the misuse of these types
of technology. The implications suggest that we are in
growing need of a system that can aid us in the recognition
of real images versus those generated by AI.
This study explores the potential of using computer
vision to enhance our newfound inability to recognise the
difference between real photographs and those that are AI-
generated. Given that there are many years worth of
photographic datasets available for image classification,
these provide examples for a model of real images.
Following the generation of a synthetic equivalent to such
data, we will then explore the output of the model before
finally implementing methods of differentiation between
the two types of image.
There are several scientific contributions with multidis-
ciplinary and social implications that arise from this study.
First, a dataset, called CIFAKE, is generated with latent dif-
fusion and released to the research community. The
CIFAKE dataset provides a contrasting set of real and fake
photographs and contains 120,000 images (60,000 images
from the exist- ing CIFAR-10 dataset (Collection of images
that are com- monly used to train machine learning and
computer vision algorithms available from:
https://www.cs.toronto.edu/ kriz/- cifar.html) and 60,000
images generated for this study), mak- ing it a valuable
resource for researchers in the field. Second, this study
proposes a method to improve our waning human ability to
recognise AI-generated images through computer vision,
using the CIFAKE dataset for classification. Finally, this
study proposes the use of Explainable AI (XAI) to further
our understanding of the complex processes involved in
synthetic image recognition, as well as visualisation of the
important features within those images. These scientific
contributions provide important steps forward in addressing
the modern challenges posed by rapid developments of
modern technology and have important implications for
ensuring the authenticity and trustworthiness of data.
The remainder of this article is as follows; the state-of-
the- art research background is initially explored in Section
II with a discussion of relevant related studies in the field.
Following this, the methodology followed by this study is
detailed in Section III, which provides the technical
implementation and the method followed for the binary
classification of real versus AI-generated imagery. The
results of these experiments are presented with discussion
in Section IV before this work is finally concluded, and
future work is proposed in Section V.

II. BACKGROUND
The ability to distinguish between real imagery and those
generated by machine learning models is important for
a number of reasons. Identification of real data provides
confirmation on the authenticity and originality of the
image; for example, a fine-tuned Stable Diffusion Model
(SDM) could be used to generate a synthetic photograph of
VOLUME 12, 15643
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
evidence of an alibi for a person who was, in reality, otherwise
elsewhere. Misinformation and fake news is a significant
modern problem, and machine-generated images could be used
to manipulate public opinion [3], [4]. Situations where synthetic
imagery is used in fake news can promote its false credibility
and have serious consequences [5]. Cybersecurity is another
major concern, with research noting that synthetically generated
human faces can be used in false acceptance attacks and have
the potential to gain unauthorised access to digital systems [6],
[7]. In [8], it was observed that synthetically generated
signatures could overcome signature verification systems with
ease.
Latent Diffusion Models are a new approach for generating
images, which use attention mechanisms and a U-Net to reverse
the process of Gaussian noise and, ultimately, use text
conditioning to generate novel images from random noise.
Details on the methodological implementation of LDM can be
found in Section III. The approach is rapidly developing but is
young, and thus literature on the subject is currently scarce. The
models are a new approach in the field of generative models;
thus, the literature is young, and few applications have been
explored. Examples of notable models include Dall-E by
OpenAI [9], Imagen from Google [10], and the open source
equivalent, SDM from StabilityAI [2]. These models have
pushed the boundaries of image quality, both in realism and
arguably in artistic ability. This has led to much debate about
the professional, social, ethical, and legal considerations of
technology [1].
The majority of research in the field is cutting-edge and
is presented as preprints and recent theses. In [11], researchers
proposed to train SDM on medical imaging data, achieving
higher-quality images that could potentially lead to increased
model abilities through data augmentations. It is worth
mentioning that in [12] and [13], diffusion models were found
to have the ability to generate audio and images. In 2021, the
results of Yi et al. [14] suggested that diffusion models were
highly capable of generating realistic artworks, fooling human
subjects into believing that the works were created by human
beings. Given this, researchers have noted that diffusion models
have a promising capacity for co- creating with human artists
[15].
DE-FAKE, proposed by Sha et al. [16], shows that images
generated by various latent diffusion approaches may contain
digital fingerprints to suggest they are synthetic. Although
visual glitches are increasingly rare given the advances in
model quality, it may be possible that computer vision
approaches will detect these attributes within images that the
human eye cannot. The Fourier transforms presented in [17]
show visual examples of these digital fingerprints.
When discussing the topic of vision, the results in [18]
suggest that optical flow techniques could detect synthetic
human faces within the FaceForensics dataset with around
81.61% accuracy. Extending to the temporal domain, [19]
proposes recurrence in AI-generated video recognition
achieving 97.1% accuracy over 80 frames due to minor visual
glitches at the pixel scale. In Wang et al. [20],

15644 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

EfficientNets and Vision Transformers are proposed within from the training dataset are used for the training of positive
a system that can detect forged images by adversarial
class ‘‘REAL’’. Therefore, 50, 000 are used for training and
models at an F1 score of 0.88 and AUC of 0.95, competing
10, 000 for testing.
with the state of the art on the DeepFake Detection
Samples of images within the CIFAR-10 dataset that
Challenge dataset while remaining efficient. In the
form the ‘‘REAL’’ class can be observed in Figure 1.
aforementioned study, a Convolutional Neural Network
was used to extract features, similarly to the approach B. SYNTHETIC DATA GENERATION
proposed in this study, prior to processing using attention-
The synthetic images generated for this study use CompVis
based approaches.
SD (https://huggingface.co/CompVis/stable-diffusion-v1-4),
Similarly, convolutional and temporal techniques were
an open source LDM. The goal is to model the diffusion of
proposed in [21] to achieve 66.26% to 91.21% accuracy in
image data through a latent space given a textual context.
a mixed set of synthetic data detection datasets.
If noise, such as that of a Gaussian distribution, is
Chrominance components CbCr within a digital image
iteratively added to an image, the image ultimately becomes
were noted in [22] as a promising route for the detection of
noise and all prior visual information is lost. To generalise,
minor pixel disparities that are sometimes present within
the reverse of this process is to, therefore, generate a
synthetic images.
synthetic image from noise. The method of reverse diffusion
Human recognition of manipulation within images is
can be put simply as, given an image x at timestep t, xt ,
wan- ing as a direct result of image generation methods
output the prediction of
improving. A study by Nightingale et al. [23] in 2017 xi−1 through the prediction of noise and subsequent removal
suggested that humans have difficulty recognising by classical means.
photographs that have been edited using image processing A noisy image xt is generated from the original x0 by the
techniques. Since this study, there has been nearly five following:
years of rapid development
in the field to date. t √ t 0 √ t
Reviewing the relevant literature has highlighted rapid x= α¯ x + 1 − α¯ ε, (1)
developments within AI-generated imagery and the chal- testing, i.e., a testing dataset of 16.6%. In this
lenges today posed with respect to its detection. Generative study, all images
models have enabled the generation of high-fidelity, photore-
alistic images within a matter of seconds that humans often
cannot distinguish between when compared to reality. This
conclusion sets the stage for the studies presented by this
work and argues the need to fill the knowledge gap when it
comes to the availability of examples of synthetic data.

III. METHOD
This section describes the methods followed by this study
in terms of their technical implementation and application
for the detection of synthetic images. This section first
describes the collection of data for the real data, and then
the methodology followed to generate the synthetic
equivalent for comparison. Sections III-A and III-B will
describe how 60,000 images are collected and 60,000
images are synthetically generated, respectively. This forms
the overall dataset of 120,000 images. Section III-C will
then describe the machine learning model engineering
approach which aims to recognise the authenticity of the
images, before Section III-D notes the approach for
Explainable AI to interpret model predictions.

A. REAL DATA COLLECTION

For the class label ‘‘REAL’’, interpreted as a positive class
value ‘‘1’’, data is collected from the CIFAR-10 dataset [24].
It is a dataset of 60, 000, 32 × 32 RGB images of real
subjects divided into ten classes. Classes within the data set
are airplane, automobile, bird, cat, deer, dog, frog, horse,
ship, and truck. There are 6, 000 images per class. For each
class, 5,000 images are used for training and 1, 000 for
VOLUME 12, 15645
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
where noise is ε, and the adjustment according to the time step t
is α¯ . The method of this study is to make use of the reverse
process of 50 noising steps, which from x50 will ultimately form
x0, the synthetic image. The neural network εθ thus
minimises the following loss function:
h i
Loss = Et,x0,ε ||ε − εθ (xt , t)||2 . (2)

Further technical details on the approach can be obtained

from [2].The model chosen for this approach is Stable Diffusion
1.4, which is trained in the LAION2B-en, LAION- high-
resolution and LAION-aesthetics v2.5 + datasets
(https://laion.ai/blog/laion-5b/). The aforementioned datasets
are a cleaned subset of the original LAION-5B dataset [25],
which contains 5.85 billion text-image pairs.
SDM is used to generate a synthetic equivalent to the
CIFAR-10 dataset which contains 6, 000 images of ten classes.
The classes are airplane, automobile, bird, cat, deer, dog, frog,
horse, ship and truck. Following observations from the CIFAR-
10 dataset, this study implements prompt modifiers to increase
diversity of the synthetic dataset, which can be observed in
Table 1. As in the real data set, 50, 000 images are used for
training data and 10, 000 for testing data, provided with a class
label to indicate that the image is not real.

C. IMAGE CLASSIFICATION
Image classification is an algorithm that predicts a class label
given an input image. The learnt features are extracted from the
image and processed in order to provide an output, in this case,
whether or not the image is real or synthetic. This subsection
describes the selected approach to classification.
In this study, the Convolutional Neural Network (CNN) [26],
[27], [28] is employed to learn from the input

15646 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

FIGURE 1. Examples of images from the CIFAR-10 image classification dataset [24].

TABLE 1. Latent diffusion prompt modifiers for generating the 10-class

synthetic dataset. All prompts are preceded by ‘‘a photograph of {a/an}’’ the weight matrix W and the bias b. The activation function
and modifiers are used equally for the 6000 images. f in this study, as in CNN, is the ReLu activation function
f (x) = max(0, x).
The goal of the network in this study is to classify
whether
the image is a real photograph or an image generated by
a LDM, and thus is a problem of binary classification.
Therefore, the output of the network is a single neuron with
the S-shaped Sigmoid activation function:
1
σ (x) = (5)
1 + e−x
The ‘‘FAKE’’ class is 0 and the ‘‘REAL’’ class is 1,
therefore, a value closer to either of the two values
represents the likelihood of that class. Although this aids in
learning, because it is differentiable, the values are rounded
to the closest value for inference.
Although the goal of the network is to use
backpropagation to reduce binary cross-entropy loss, this
study also notes an extended number of classification
images. It is the concatenation of two main networks with metrics. These are the Precision, which is a measure of how
intermediate operations. These are the convolutional layers many of the predictive positive cases are positive, a
and the fully connected layers. The initial convolutional metric which allows for the
analysis of false-positives:
network within the overall model is the CNN, which can be True positives
Precision = True positives + False . (6)
operationally generalised for an image of dimensions x and
a positives
filter matrix w as follows: The Recall which is a measure of how many positive cases
are
M N correctly predicted, which enables analysis of false-negative
XX
(x ∗ w)(i, j) = x(i + m − 1, j + n − 1)w(m, predictions:
True positives ,
n), (3) Recall = (7)
True positives + False negatives
m=1 n=1
where (i, j) is the output for the feature map, and (m,
n)
represents the location of the filter w. The output is derived Then, a pooling operation is performed to reduce the
by applying convolutional operations to the input x with spatial dimensions and flatten the output so it can be
each of the filters (which are learnable) and applying an entered into densely connected layers. For L = HWD
activation function f , which, in the context of this study, is (height, width, and dimensions), the flattened one-
dimensional output vector is
the Rectified
simply x = [x1, x2, . . . , xL]. The output vector y is
Linear Unit (ReLu) f (x) = max(0, x). ultimately the output from the dense layer(s) as y =
For an image of (height, width) dimensions and fil- f (WL + b), for
ters depending on the filter kernel of (heightkernel) and
(widthkernel) with a stride = 1 and no padding for simplicity,
the output would have dimensions:
(height − heightkernel + 1, width − widthkernel + 1).
(4)
VOLUME 12, 15647
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
This measure is particularly important in this case, as it is in
fraud detection, since a false negative would falsely accuse the
author of generating their image with AI. Finally, the F-1 score
is considered:
Precision × Recall
F1 score = 2 × ,
Precision + Recall (8)
which is a unified metric of precision and recall.
The dataset that forms the classification is the collection of
real images and the equivalent synthetic images gen- erated,
detailed in Sections III-A and III-B, respectively. 100, 000
images are used for training (50, 000 real images
and 50, 000 synthetic images), and 20, 000 are used for
testing (10, 000 real and 10, 000 synthetic).
Initially, CNN architectures are benchmarked as a lone
feature extractor. That is, the filters of {16, 32, 64, 128} are
benchmarked in layers of {1, 2, 3}, flattened, and

15648 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

FIGURE 2. Examples of AI-generated images within the dataset contributed by this study, selected at random with regards to their real
CIFAR-10 equivalent labels.

connected directly to the output neurone. The topology of P P

where αk is the global average pooling 1
∂y c
of
the highest performing feature extractor is then used to Z i j ∂A k
i,j
compare the highest performing dense network featuring ∂y c
spatial locations Z , k are the gradients of the model.
∂A i,j
and
{32, 64, 128, 256, . . . , 4096} rectified linear units in The approach is used for interpretation in the final step
layers of this study, given the random data selected from the two
of {1, 2, 3}. These 36 artificial neural networks are classes. Due to the nature of heatmapping, the results of the
then compared with regard to classification metrics to algorithm are visually interpreted with discussion.
derive the
topology that performs best. E. EXPERIMENTAL HARDWARE AND SOFTWARE
The neural networks used for the detection of AI-generated
D. EXPLAINABLE AI images were engineered with the TensorFlow library [31].
While deep learning approaches often lead to impressive All TensorFlow seeds were set to 1 for replicability. The
predictive ability, many algorithms are black boxes that Latent Diffusion model used for the generation of synthetic
provide no reasoning for their classification. The aim of data was Stable Diffusion version 1.4 [2]; Random seed
Explainable AI (XAI) is to extract meaning from vectors were denoised for a total of 50 steps to form images
algorithms and provide readable interpretations of why a and the Euler Ancestral scheduler was used. Synthetic
prediction or decision is being made [29]. Regarding the images were rendered at a resolution of 512px before
experiments in this work, the CNN simply predicts that an resizing to 32px by bilinear interpolation to match the
image is real or synthetic, and then XAI is used to provide resolution of CIFAR-10.
interpretations as to why the image is real or synthetic. All algorithms in this study were executed using a single
Given that the literature shows that humans have a major Nvidia RTX 3080Ti GPU, which has 10,240 CUDA cores,
difficulty in recognising synthetic imagery, it is important a clock speed of 1.67 GHz, and 12GB GDDR6X VRAM.
to display and visualise minor defects within the image that
could suggest that it is not real. IV. RESULTS AND OBSERVATIONS
The method selected for explainable AI (XAI) and This section presents examples of the dataset followed by
interpretation is Gradient Class Activation Mapping (Grad- the findings of the planned computer vision experiments.
CAM) [30]. Grad-CAM interprets the gradients of the The dataset is also released to the public research
predicted class along with the CNN feature maps, which community for use in future studies, given the important
can therefore be spatially localised with respect to the implications of detecting AI-generated imagery.1
original
input (the image) and produce a heatmap. This is generated
by the Recitified Linear Unit (ReLU) function as:
VOLUME 12, 15649
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
c Identification
X
LGrad = ReLU ( αk Ak ), (9) 1
The Dataset can be downloaded from:
−CAM
k https://www.kaggle.com/datasets/ birdy654/cifake-real-and-ai-generated-
synthetic-images

15650 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

TABLE 3. Observed validation loss for the filters within the

convolutional neural network.

FIGURE 3. Examples of visual defects found within the synthetic TABLE 4. Observed validation precision for the filters within
image dataset. the convolutional neural network.

TABLE 2. Observed classification accuracy metrics for feature extraction

networks.

TABLE 5. Observed validation recall for the filters within the

convolutional neural network.

A. DATASET EXPLORATION
Random samples of images used in this study and within
the dataset provided can be observed in Figure 2. Five
images are presented for each class label, and all of the
images within this figure are synthetic, which have been
generated by the SDM. Note within this sample that the
images are high- quality and, for the most part, seem to be TABLE 6. Observed validation F1-Score for the filters within
difficult to discern as synthetic by the human eye. Synthetic the convolutional neural network.
photographs are representative of their counterparts from
reality and feature complex attributes such as depth of field,
reflections, and motion blur.
It can also be observed that there are visual imperfections
within some of the images. Figure 3 shows a number of
examples of the win of the dataset in which the model has
output images with visual glitches. Given that the LAION
dataset provides physical descriptions of the image content,
faced by the CNN is that of binary classification, whether or
little to no information on text is provided, and thus it can
not the image is a real photograph or the output of an LDM.
be seen that the model produces shapes similar to
The validation accuracy of the results and the loss metrics
alphabetic characters. Also observed here is a lack of
for the feature extractors can be found in Tables 2 and 3,
important detail, such as the case of a jet aircraft that has no
respectively. All feature extractors scored relatively well
cockpit window. It seems that this image has been produced
without the need for dense layers to process feature maps,
by combining the knowledge of jet aircraft (in particular,
with an average classification accuracy of 91.79%. The
the engines) along with the concept of an Unmanned Aerial
lowest loss feature extractor was found to use two layers of
Vehicle’s chassis. Finally, there are also some cases of
32 filters, which led to an overall classification accuracy of
anatomical errors for living creatures, seen in these
92.93% and a binary cross-entropy loss of 0.18. The highest
examples through the cat’s limbs and eyes.
accuracy model, two layers of 128 filters, scored 92.98%
Complex visual concepts are present within much of the
with
dataset, with examples shown in Figure 4. Observe that
a loss of 0.221.
the ripples in the water and reflections of the entities are
Extended validation metrics are presented in Tables 4, 5,
highly realistic and match what would be expected within
and 6, which detail validation precision, recall, and F1
a photograph. In addition to complex lighting, there is also
scores, respectively. The F1 score, which is a unification of
evidence of depth of field and photographic framing.
precision and recall, had a mean value of 0.929 with the
highest being 0.936. A small standard deviation of 0.003
B. CLASSIFICATION RESULTS
was observed.
In this subsection, we present the results for the computer
Following these experiments, the lowest-loss feature
vision experiments for image classification. The problem
extractor is selected for further engineering of the network
topology. This was the model that had two layers of
VOLUME 12, 15651
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
32 convolutional filters.

15652 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

FIGURE 4. A selection of AI-generated images within the dataset. Examples of complex visual attributes
generated by the latent diffusion model that include realistic water and reflections.

TABLE 7. Observed validation accuracy for the dense layers within the
convolutional neural network. TABLE 9. Observed validation precision for the dense layers within
the convolutional neural network.

TABLE 8. Observed validation loss for the dense layers within

the convolutional neural network. TABLE 10. Observed validation recall for the dense layers within
the convolutional neural network.

The results of the general network engineering are of the network topology.
presented in Tables 7 and 8, which contain the validation
accuracy and loss, respectively. The lowest loss observed
was
0.177 binary cross-entropy when the CNN was followed by
three layers of 64 rectified linear units. The highest
accuracy, on the other hand, was 93.55%, which was
achieved by implementing a single layer of 64 neurons.
Additional validation metrics for precision, recall, and F-
1 score are also provided in Tables 9, 10, and 11,
respectively. Similarly to the prior experiments, the
standard deviation of F1-scores was relatively low at 0.003.
The highest F-1 score was the network that used a single
dense layer of 64 rectified linear units, with a value of
0.936. The aforementioned highest F1 score model is
graphically detailed in Figure 5 to provide a visual example
VOLUME 12, 15653
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
TABLE 11. Observed validation F1-Score for the dense layers within the
convolutional neural network.

Figure 6 shows examples of the interpretation of predic- tions

via Grad-CAM. Brighter pixels in the image represent

15654 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

FIGURE 5. An example of one of the final model architectures following hyperparameter search for the classification of
real or AI-generated images.

and featured complex visual attributes, and that binary

classification could be achieved with around 92.98% accu-
racy. Grad-CAM interpretation revealed interesting
concepts within the images that were useful for predictions.
In addition to the method proposed in this study, a
signifi- cant contribution is made through the release of the
CIFAKE dataset. The dataset contains a total of 120,
000 images (60, 000 real images from CIFAR-10 and
60,000 synthetic images generated for this study). The
FIGURE 6. Gradient class activation maps (Grad-CAM) overlays and raw CIFAKE dataset provides the research community with a
heatmaps for prediction interpretation. Top examples show real images
and bottom examples show AI-generated images. Brighter pixels valuable resource for future work on the social problems
represent features contributing to the output class label. faced by AI-generated imagery. The dataset provides a
significant expansion of the resource availability for the
areas that contribute more to the decision of the CNN. It development and testing of applied computer vision
can be observed that there is a significantly different approaches to this problem.
distribution of features given the binary classification The reality of AI generating images that are
problem. Firstly, the classification of real images can be indistinguish- able from real-life photographic images
interpreted as a more holistic approach in which the raises fundamental questions about the limits of human
majority of the contents of the image are useful for perception, and thus this study proposed to enhance that
recognition. However, the classification of synthetic images ability by fighting fire with fire. The proposed approach
is somewhat more atomistic and sparse. Note that for the addresses the challenges of ensuring the authenticity and
recognition of AI-generated imagery, activation occurs in trustworthiness of visual data.
select parts of the image that are more likely to present Future work could involve exploring other techniques to
visual glitches that are difficult to recognise with the classify the provided dataset. For example, the implemen-
human eye. An example of this can be seen for the tation of attention-based approaches is a promising new
image of the frog, where an out-of-focus bokeh is the only field that could provide increased ability and an alternative
attribute that suggests the image is not real. For the truck, method of explainable AI. Furthermore, with even further
only the radiator grill pattern is considered useful for improvements to synthetic imagery in the future, it is
classification. important to consider updating the dataset with images
The XAI approach also shows an interesting mechanic in generated by these approaches. Furthermore, considering
a more general sense. Given the examples of airplane, bird, generating images from other domains, such as human
frog, horse, and ship, note that the object within the image faces and clinical scans, would provide additional datasets
has little to no class activation overlay whatsoever. This for this type of study and expand the applicability of our
suggests that the actual focus of the image itself, the entity, proposed approach to other fields of research.
contains almost no useful features for synthetic image Finally, in conclusion, this study provides contributions
recognition. This suggests that the model is often available to the ongoing implications of AI-generated images. The
to produce a near-perfect representation of the entity. pro- posed approach supports important implications of
ensuring data authenticity and trustworthiness, providing
V. CONCLUSION AND FUTURE WORK not only a system that can recognise synthetic images, but
This study has proposed a method to improve our waning also data and interpretation. The public release of the
ability to recognise AI-generated images through the use CIFAKE dataset generated within this study provides a
of Computer Vision and to provide insight into predictions valuable resource for interdisciplinary research.
with visual cues. To achieve this, this study proposed the
generation of a synthetic dataset with Latent Diffusion, VI. AVAILABILITY OF DATA AND MATERIALS
recognition with Convolutional Neural Networks, and inter- The datasets generated and analysed during the cur-
pretation through Gradient Class Activation Mapping. The rent study are available in the CIFAKE repository,
results showed that the synthetic images were high quality https://www.kaggle.com/datasets/birdy654/cifake-real-and-
VOLUME 12, 15655
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
ai-generated-synthetic-images.

15656 VOLUME 12,

2024
J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification

REFERENCES tiny images,’’ 2009.

[1] K. Roose, ‘‘An AI-generated picture won an art prize. Artists aren’t
happy,’’
New York Times, vol. 2, p. 2022, Sep. 2022.
[2] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer,
‘‘High-resolution image synthesis with latent diffusion models,’’ in Proc.
IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2022,
pp. 10684–10695.
[3] G. Pennycook and D. G. Rand, ‘‘The psychology of fake news,’’ Trends
Cogn. Sci., vol. 25, no. 5, pp. 388–402, May 2021.
[4] B. Singh and D. K. Sharma, ‘‘Predicting image credibility in fake news
over social media using multi-modal approach,’’ Neural Comput. Appl.,
vol. 34, no. 24, pp. 21503–21517, Dec. 2022.
[5] N. Bonettini, P. Bestagini, S. Milani, and S. Tubaro, ‘‘On the use of
Benford’s law to detect GAN-generated images,’’ in Proc. 25th Int. Conf.
Pattern Recognit. (ICPR), Jan. 2021, pp. 5495–5502.
[6] D. Deb, J. Zhang, and A. K. Jain, ‘‘AdvFaces: Adversarial face
synthesis,’’ in Proc. IEEE Int. Joint Conf. Biometrics (IJCB), Sep. 2020,
pp. 1–10.
[7] M. Khosravy, K. Nakamura, Y. Hirose, N. Nitta, and N. Babaguchi,
‘‘Model inversion attack: Analysis under gray-box scenario on deep
learning based face recognition system,’’ KSII Trans. Internet Inf. Syst.,
vol. 15, no. 3, pp. 1100–1118, Mar. 2021.
[8] J. J. Bird, A. Naser, and A. Lotfi, ‘‘Writer-independent signature
verification; evaluation of robotic and generative adversarial attacks,’’
Inf. Sci., vol. 633, pp. 170–181, Jul. 2023.
[9] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen,
and I. Sutskever, ‘‘Zero-shot text-to-image generation,’’ in Proc. Int.
Conf. Mach. Learn., 2021, pp. 8821–8831.
[10] C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton,
S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes,
T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi, ‘‘Photorealistic text-
to-image diffusion models with deep language understanding,’’ 2022,
arXiv:2205.11487.
[11] P. Chambon, C. Bluethgen, C. P. Langlotz, and A. Chaudhari, ‘‘Adapting
pretrained vision-language foundational models to medical imaging
domains,’’ 2022, arXiv:2210.04133.
[12] F. Schneider, O. Kamal, Z. Jin, and B. Schölkopf, ‘‘Moûsai: Text-to-
music generation with long-context latent diffusion,’’ 2023,
arXiv:2301.11757.
[13] F. Schneider, ‘‘ArchiSound: Audio generation with diffusion,’’ M.S. thesis,
ETH Zurich, Zürich, Switzerland, 2023.
[14] D. Yi, C. Guo, and T. Bai, ‘‘Exploring painting synthesis with diffusion
models,’’ in Proc. IEEE 1st Int. Conf. Digit. Twins Parallel Intell.
(DTPI), Jul. 2021, pp. 332–335.
[15] C. Guo, Y. Dou, T. Bai, X. Dai, C. Wang, and Y. Wen, ‘‘ArtVerse:
A paradigm for parallel human–machine collaborative painting creation
in Metaverses,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 53, no. 4,
pp. 2200–2208, Apr. 2023.
[16] Z. Sha, Z. Li, N. Yu, and Y. Zhang, ‘‘DE-FAKE: Detection and
attribution of fake images generated by text-to-image generation
models,’’ 2022, arXiv:2210.06998.
[17] R. Corvi, D. Cozzolino, G. Zingarini, G. Poggi, K. Nagano, and
L. Verdoliva, ‘‘On the detection of synthetic images generated by
diffusion models,’’ 2022, arXiv:2211.00680.
[18] I. Amerini, L. Galteri, R. Caldelli, and A. Del Bimbo, ‘‘Deepfake video
detection through optical flow based CNN,’’ in Proc. IEEE/CVF Int.
Conf. Comput. Vis. Workshop (ICCVW), Oct. 2019, pp. 1205–1207.
[19] D. Güera and E. J. Delp, ‘‘Deepfake video detection using recurrent
neural networks,’’ in Proc. 15th IEEE Int. Conf. Adv. Video Signal Based
Surveill. (AVSS), Nov. 2018, pp. 1–6.
[20] J. Wang, Z. Wu, W. Ouyang, X. Han, J. Chen, Y.-G. Jiang, and S.-N. Li,
‘‘M2TR: Multi-modal multi-scale transformers for Deepfake detection,’’
in Proc. Int. Conf. Multimedia Retr., Jun. 2022, pp. 615–623.
[21] P. Saikia, D. Dholaria, P. Yadav, V. Patel, and M. Roy, ‘‘A hybrid CNN-
LSTM model for video Deepfake detection by leveraging optical flow
features,’’ in Proc. Int. Joint Conf. Neural Netw. (IJCNN), Jul. 2022,
pp. 1–7.
[22] H. Li, B. Li, S. Tan, and J. Huang, ‘‘Identification of deep network
generated images using disparities in color components,’’ Signal
Process., vol. 174, Sep. 2020, Art. no. 107616.
[23] S. J. Nightingale, K. A. Wade, and D. G. Watson, ‘‘Can people identify
original and manipulated photos of real-world scenes?’’ Cognit. Res.,
Princ. Implications, vol. 2, no. 1, pp. 1–21, Dec. 2017.
[24] A. Krizhevsky and G. Hinton, ‘‘Learning multiple layers of features from
VOLUME 12, 15657
2024
 J. J. Bird, A. Lotfi: CIFAKE: Image Classification and Explainable
Identification
[25] C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman,
M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman,
P. Schramowski, S. Kundurthy, K. Crowson, L. Schmidt, R. Kaczmarczyk,
and J. Jitsev, ‘‘LAION-5B: An open large-scale dataset for training next
generation image-text models,’’ 2022, arXiv:2210.08402.
[26] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521, no.
7553, pp. 436–444, 2015.
[27] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang,
G. Wang, J. Cai, and T. Chen, ‘‘Recent advances in convolutional neural
networks,’’ Pattern Recognit., vol. 77, pp. 354–377, May 2018.
[28] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, ‘‘A survey of convolutional
neural networks: Analysis, applications, and prospects,’’ IEEE Trans. Neural
Netw. Learn. Syst., vol. 33, no. 12, pp. 6999–7019, Dec. 2022.
[29] D. Gunning, M. Stefik, J. Choi, T. Miller, S. Stumpf, and G.-Z. Yang, ‘‘XAI—
Explainable artificial intelligence,’’ Sci. Robot., vol. 4, no. 37, Dec. 2019, Art.
no. eaay7120.
[30] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and
D. Batra, ‘‘Grad-CAM: Visual explanations from deep networks via gradient-
based localization,’’ in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), Oct. 2017,
pp. 618–626.
[31] M. Abadi et al. (2015). TensorFlow: Large-Scale Machine Learning on
Heterogeneous Systems. [Online]. Available: https://www.google.com/ search?
q=httpsSoftware+available+from+tensorflow.org&rlz=1C1GCEU
_enIN1087IN1087&oq=httpsSoftware+available+from+tensorflow.org
&gs_lcrp=EgZjaHJvbWUyBggAEEUYOTIKCAEQABiABBiiBDIKCA
IQABiiBBiJBTIKCAMQABiABBiiBDIKCAQQABiABBiiBDIKCAUQ
ABiABBiiBNIBCDY0MDZqMGo5qAIAsAIA&sourceid=chrome&ie
=UTF-8

JORDAN J. BIRD received the B.Sc. and Ph.D.

degrees in computer science from Aston Uni- versity,
U.K. He is currently a Senior Lecturer of computer
science with Nottingham Trent Uni- versity, U.K. He
received significant external grant funding toward his
research projects, which involve applications of
artificial intelligence in the real-world. His research
interests include artificial intelligence (AI), focusing
on human–robot inter- action (HRI), machine
learning (ML), deep learn-
ing, transfer learning, and data augmentation. His professional academic
contributions include roles as a technical program committee member and the
deep learning session chair of several international conferences.

AHMAD LOTFI (Senior Member, IEEE) received the

B.Sc. degree in control systems from the Isfahan
University of Technology, Iran, the M.Tech. degree in
control systems from the Indian Institute of
Technology Delhi, India, and the Ph.D. degree in
learning fuzzy systems from The University of
Queensland, Australia, in 1995. He is currently a
Professor of computational intelligence with
Nottingham Trent University, where he is leading
the research group in computational
intelligence and applications. He has authored and coauthored over 200
scientific papers in the area of computational intelligence, the Internet of Things,
abnormal behavior recognition, and ambient intelligence in highly prestigious
journals and international conferences. He received external funding from
Innovate UK, EU, and industrial companies to support his research. His
research interest includes the identification of progressive changes in behavior
of older individuals suffering from dementia or any other cognitive
impairments. Accurate identification of progressive changes through utilization
of unobtrusive sensor network or robotics platform will enable carers (formal
and informal) to intervene when deemed necessary. Research collaboration is
established with world-leading researchers. He has involved in collaboration
with many healthcare commercial organizations and end-users. He has been
invited as an expert evaluator and a panel member of many European and
international research programs.

15658 VOLUME 12,

2024