IDS 705 Final Report

Detecting AI-Generated Fake Images

Towards a Generalizable Fake Facial Image Detector via Neural Networks

Team Members:
Kabber, Pooja
Lai, Scott
Oh, Song Young
Wang, Emma

Team Number: 5
Detecting AI-Generated Fake Images

Abstract
With the steady rise in innovative new deep learning techniques, the detection of AI-generated fake
images has become a critical research problem. In this project, we aim to apply a convolutional neural
network based model to detect a particular kind of fakes - facial images. This idea was developed from
previous studies done in this space, including Sabithaa et al. (2020) and Zhang et al. (2020). We use
the current state-of-the-art pre-trained ResNet model to address vanishing gradients in deep neural
networks, as proposed by He et al. (2016). We also run experiments using the light-weight pre-trained
MobileNetV2. Our models are fine-tuned and evaluated using the FFHQ (Flickr-Faces-HQ) dataset as
the benchmark. Through our experimentation, we achieve high accuracy in detecting fake images, and
analyze the impact of different categories on model performance, obtaining an overall accuracy of 80%
with MobileNetV2 and an accuracy of 89% with ResNet50. Our findings can be applied to real-world
applications where fake image detection is critical.

Introduction
Image-generation AI models have made significant progress in recent years, particularly with the
advances in neural network techniques. This has led to the development of several impressive models
including DALL-E 2 developed by OpenAI. DALL-E 2 is a transformer autoregressive model in which
the model learns to map the input text to a latent space representation, which is then used to generate
a corresponding image. Thus, DALL-E 2 can be effectively used for fake image generation through
textual descriptions. DALL-E 2's ability to generate high-quality images has significant implications for a
wide range of applications, including entertainment, advertising, digital art, and even architecture
design.

However, the use of AI-generated images raises important questions about the ethical implications of
using such images. For example, AI-generated images can be misused to create fake images or videos
of individuals, which may potentially lead to serious consequences such as identity theft or reputational
harm. For this reason, this project aims to build a machine learning model that can differentiate
between AI-generated images created by DALL-E 2 and real human face images. Our project is driven
by the potential ethical implications of using AI-generated images in various contexts, such as deep
fakes and racial bias. By uncovering valuable insights into these differences, we hope to contribute to
this significant area of study and promote ethical practices in the use of AI-generated images.

Background
The detection of AI-generated images is crucial due to the potential for their use in various malicious
activities, including deep fake videos, synthetic identity fraud, propaganda, and art fraud. However, the
increasing sophistication and realism of AI-generated images have made it challenging for humans to

1
Detecting AI-Generated Fake Images

distinguish between real and fake images. To explore the potential for fraud detection in images, we are
using DALLE-2 to generate fake images from our real image dataset in our study, given its
advancement in generating images that are difficult to detect as fraud by human eyes.

Deep fake videos are a compelling example of the misuse of AI-generated images. These videos
manipulate existing footage using AI technology to replace the face of one person with another,
creating fake news, spreading propaganda, and damaging the reputation of individuals. Similarly,
AI-generated images can also be used for synthetic identity fraud, where criminals create fake identities
using stolen personal information and AI-generated images to commit fraudulent activities like opening
bank accounts or applying for loans. Additionally, the recent controversy surrounding AI-generated
artwork in Taiwan has highlighted the potential for fraud in art sales. This debate centers on the use of
computer-generated tools to create art and the concerns that this undermines the work of traditional
artists who rely solely on their own skills and expertise. As AI-generated images become increasingly
sophisticated, it is becoming more challenging for the public to differentiate between real and fake
artwork, creating opportunities for fraud.

Our research aims to develop a machine learning pipeline that can detect fraud in images, thereby
safeguarding against the negative consequences of AI-generated images. By gaining a deeper
understanding of the differences between AI-generated facial images and real human faces, we hope to
contribute to the development of more responsible and ethical practices for using AI-generated images
across different fields.

In the field of computer vision research, there has been a surge of publications related to DALL-E 2
over the past few months. These papers have garnered an increasing amount of attention due to the
model’s impressive capabilities and its potential applications in various fields. In one recent study, Sha
et al. (2022) proposed a novel method for detecting and attributing fake images. The authors conducted
extensive experiments on four popular text-to-image generation models, namely DALL-E 2, Stable
Diffusion, GLIDE, and Latent Diffusion, using two benchmark prompt-image datasets. The analysis
results revealed that fake images generated by various models can be distinguished from real ones
based on a common artifact shared by fake images from different models.

Another paper by Ojha et al. (2023) addressed the problem of training neural networks to detect fake
images and evaluated recent text-to-image generation models, including Latent Diffusion, GLIDE, and
DALL-E. The results suggested that there is something common between fake images generated from
a GAN and those from a diffusion model, although the specific similarity remains an open question.
Meanwhile, Corvi et al. (2022) studied the performance of current detectors, developed for
GAN-generated images, on twelve kinds of new synthetic images, including DALL-E 2 images. The
study particularly focused on challenging social-networks scenarios involving image compression and
resizing.

2
Detecting AI-Generated Fake Images

Title Model Used Accuracy Architecture Dataset Code/Source

Deep convolutional 96.45% The model consists Real: FFHQ Sabithaa et al.
Enhanced model neural network (CNN) of multiple (2020)
for fake image designed to detect fake convolutional and
detection (EMFID) images using a max pooling layers,
using combination of histogram followed by fully
convolutional and wavelet-based connected layers
neural networks feature extraction. and a softmax
with histogram and output layer that
wavelet based produces a
feature extractions probability
distribution over the
two classes (fake
and real).

Deep residual ResNet The authors The authors ImageNet dataset https://github.com/p
Learning for Image evaluated experimented with https://www.image-n ytorch/vision/blob/m
Recognition ResNet on the various depths of et.org/download.php ain/torchvision/mode
ImageNet ResNet, from 18 ls/resnet.py
dataset and layers to 152
achieved layers, and found
state-of-the–art that increasing the
performance, depth beyond a
with a top-5 error certain point led to
rate of 3.57% for diminishing returns
ResNet-152. in performance.

DE-FAKE: ResNet18 0.871 Their 18-layer - Real: MSCOCO Kaiming He, et al.
Detection ResNet has no extra (20,000 images) (2016)
and Attribution parameter
of Fake Images compared to its - Fake: Generated
Generated by plain counterpart. 20,000 images by
Text-to-Image using the prompts of
Generation real images through
Models Stable Diffusion

On the Detection ResNet50 0.595 block = BottleNeck ProGAN images S.-Y. Wang, et al.
of Synthetic layers = [3,4,6,3] (2020)
Images Generated num_classes = https://github.com/p
by 1000 eterwang512/CNND
Diffusion Models zero_init_residual = etection
False

Towards Universal ResNet50 0.8425 They vary the voting ProGAN images https://github.com/Y
Fake Image pool size from k=1 uheng-Li/UniversalF
Detectors that to k=9. More details akeDetect
Generalize Across will be released (Data has not been
Generative Models soon. released yet)

Detecting ResNet50 0.941 They transfer the Public dataset N/A

Computer weights of layers comprising of 9,700
Generated Images pre-trained on images
with Deep ImageNet data to
Convolutional their deep model,
Neural Networks using a transfer
learning approach.

Distinguishing CNN 0.9337 Each of five - Real: RAISE N/A

Computer-Generat (50 epochs) convolutional layers dataset
ed Graphics from is followed by a (1,800 images)
Natural Images 0.9312 batch normalization
Based on Sensor (80 epochs) layer, a ReLU layer, - Fake: Level-design
Pattern Noise and and an average reference database
Deep Learning pooling layer. (1,800 images)

3
Detecting AI-Generated Fake Images

Multi-attentional EfficientNet-b4 0.976 in_channels = 3 - FaceForensics++ https://github.com/y

Deepfake (high-quality) out_channels = 32 - DFDC octta/multiple-attenti
Detection on
0.8869
(low-quality)

Face XceptionNet 0.701 They replace the FaceForensics++ Francois Chollet

Forensics++: final FC layer with (2017)
Learning to Detect two outputs and the
Manipulated Facial other layers are
Images initialized with the
ImageNet weights.

Table 1. Comparison of Related Work

Data
The dataset we use for our experiments consists of two classes: real and fake, the latter of which is
generated by Dall-E 2 from the auto-generation pipeline we build. For the real dataset, we adopt the
Flickr-Faces-HQ (FFHQ) dataset, a high-quality collection of 70,000 PNG images at 1024×1024
resolution. These images were crawled from their google drive to Amazon S3, a cloud-based storage
service provided by AWS. This dataset comprises real human images with considerable variation in
terms of age, ethnicity, and image background, and it also includes a broad coverage of accessories
such as eyeglasses, sunglasses, hats, and more. Since our data was originally unlabeled with respect
to these categories and human annotation is time-consuming, we annotate our images using machine
learning models in cloud-based development environments such as GitHub Codespaces and Google
Colab. Meanwhile, to ensure fair comparisons, the fake images in our dataset were also set to a
resolution of 1024×1024, matching the resolution of the real images. We also generate corresponding
fake images for our set of real images to ensure that the learning of the model is to differentiate
between real and fake images and not to separate the images using some other criteria.

Data Annotation

To ensure that our dataset covers a diverse range of human faces, we employed four pre-trained
classification models for the following categories: facial expression (Trpakov, 2022), race (Parkhi et al.),
gender (Parkhi et al.), and glasses (Mandgi, 2021). We obtained the facial expression model from
HuggingFace, which recognizes seven distinct emotions. For more convenient analysis and adequate
representation of all classes, we converted the seven expressions to the positive, negative and neutral
scale. For glasses detection, we utilized a binary classification approach that involved cropping each
image based on facial landmark coordinates for the nasal area and converting the image into black and
white to identify the presence of a strip of black pixels (0s) on the face. For race and gender, we used
the VGG-Face model pre-trained on ImageNet. The race categories were asian, white, middle eastern,
indian, latino and black. We absorbed middle eastern and indian into asian since they were not
represented adequately individually. Gender has two categories: male and female.

These labeling results were then aggregated into a single data frame, with one column for the image
file name and four columns for the labeling results. This rigorous process allowed us to create a more

4
Detecting AI-Generated Fake Images

representative dataset by selecting a sample of sample size 21k from the 70k real images such that
each category and class were adequately represented and in a comparable ratio to each other. The
final ratios for the categories and classes are as follows:

Facial Positive 0.52

Expression
Negative 0.25

Neutral 0.23

Glasses Glasses 0.67

No Glasses 0.33

Gender Man 0.59

Woman 0.41

Race White 0.31

Asian 0.3

Latino Hispanic 0.3

Black 0.11

Table 2. Ratios of classes for all facial categories

Another reason for selecting these four facial characteristics is because we aim to analyze if images
belonging to these categories may be more challenging for DALLE-2 to generate vivid images, thereby
making it easier for our model to detect fakes, for example, a face with eyeglasses. Our objective is to
investigate whether other facial features would impact its performance (gender, glasses, etc.).
Nonetheless, it is important to note that our annotations only cover some types of human faces and do
not represent all categories.

During the annotation process, we manually labeled a few hundred images to estimate the performance
of the models. We observed that the facial expression model had an accuracy of approximately 70%.
Meanwhile, the models for detecting faces wearing glasses and classifying gender were more accurate,
with an accuracy of 91%. Additionally, the model for race classification had an accuracy of 86% and
around 96% for gender. Although some images were misclassified, we still wanted to maintain the
machine labeling results due to limited time and the primary aim of our data annotation, which was to
ensure we have enough images in many facial categories. Nevertheless, we acknowledge that our
annotations may not be completely accurate and this can affect the efficacy of our analysis of the
detector’s performance.

5
Detecting AI-Generated Fake Images

Figure 1. Characteristics for Real Image vs Fake Image. The first column show the four class that we use to do
the picture classification: Gender, Glasses, Race and Facial Expression; Second column contains the sub-classes
for each class; third column contains the real image we use for AI image generation; the last column is the fake
images generate from DallE-2.

6
Detecting AI-Generated Fake Images

Pipeline of Preprocessing

Figure 2. Workflow diagram

To ensure a balanced dataset, we performed four separate classification analyses on the original
dataset of 70,000 images. The images were classified into classes for four categories: glasses, facial
expressions, race, and gender. Using the number of images in each category, we randomly selected a
total of 21,000 images from across all four categories based on the class weight for each class. This
approach allowed us to create a well-balanced dataset for training our models.

The AI images were generated using OpenAI's DALL-E 2 model, ensuring consistency with the real
images in our dataset. We believe that this may allow us to ensure that the learning of the model is in
fact the difference between real and fake images and not differences in other characteristics. To
facilitate efficient model training, we utilized both Google Colab and CodeSpaces with AWS S3 to run
multiple models, and stored both the real and fake images in Google Drive and the AWS S3 bucket. We
also made the data publicly available in Google Drive, to be used by the community for other research.

After training the models, we saved them on Hugging Face, a popular platform for sharing and
deploying NLP and ML models. To enable easy access and utilization of our models, we built a web

7
Detecting AI-Generated Fake Images

application using Streamlit and Hugging Face. This application allows users to detect fake images and
classify them into the four categories we identified during the selection process.

In conclusion, our preprocessing pipeline was carefully designed to ensure a balanced dataset and
efficient model training. We believe that open access to data and models is crucial for advancing
research and enhancing model accuracy. Therefore, we are excited to make our dataset and models
available to the research community to facilitate further studies in this field.

Experiments
Project Architecture

Our project aims to build a robust detector of fake images with the following workflow. First, we acquire
70,000 images from the FFHQ dataset and formulate our dataset which is stored in AWS S3. We then
annotate the 70,000 images, including attributes such as facial expression, wearing glasses, races, and
gender. We select 21k real images and build an image-to-image pipeline that generates a fake dataset
of 21k images. We then perform image preprocessing (ensuring RGB color model and resizing the
images to 180x180 for input to the model), and feature extraction, followed by a model training process.
The classification will be done using two types of neural network models MobileNetV2 and ResNet50,
and the results will be evaluated.

Methodology

Our work is intended to build a neural network model to detect fake AI images, based on previous work
done including Sabithaa et al. (2020), Zhang et al. (2020), and He et al. (2016). Sabithaa et al. (2020)
used an enhanced model for detecting fake images generated by artificial intelligence using
convolutional neural networks with histogram and wavelet-based feature extractions. In our study, the
real dataset we used is the same as one of their evaluation datasets, FFHQ dataset, which would be a
good reference for our model building process. Zhang et al. (2020) also adopted a CNN-based
architecture for detecting fake images. We particularly paid attention to He et al. (2016) which proposed
the ResNet model developed for addressing the problem of vanishing gradients in deep neural
networks. Our study can fine-tune their pre-trained model and visualize outputs to gain insights into the
model’s predictions. This visualization would be highly useful to our study, because we aim to achieve
high accuracy while understanding the factors influencing detection performance. Based on this
research, we will be using ResNet50. We also chose to use MobileNetV2 since it is a lightweight
version of ResNet50, and to explore its performance as compared to ResNet50.

MobileNetV2 (Visualization displayed in Appendix 1.)

8
Detecting AI-Generated Fake Images

MobileNetV2 model, as a pre-trained base model for transfer learning, is a deep learning architecture
that is designed for mobile devices and has a small computational footprint. The model was trained on
the ImageNet dataset, which has over 1 million images with 1000 different classes. The model has
been shown to achieve high accuracy on image classification tasks while requiring fewer parameters
and computation compared to other popular architectures. We adopted the pre-trained MobileNetV2
model with its weights trained on the ImageNet dataset, and the top layer of the model is excluded by
setting include_top to False. This allows us to reuse the feature extraction capabilities of the model
while adding our custom classification layer on top.
During the training process, we freeze all layers in the base model, so only the newly added
classification layers are trainable during training. The added layers include a data augmentation layer to
artificially increase the dataset's size and prevent overfitting, the base MobileNetV2 model, followed by
a Global Average Pooling layer, and three fully connected layers with ReLU activation, Dropout layers,
and L2 regularization. The final layer has a sigmoid activation function since this is a binary
classification task.

The model is compiled with the Adam optimizer, binary cross-entropy loss function, and accuracy
metric. The goal is to train the model to classify images into two classes, fake and real, with high
accuracy, more specifically high score in precision of fake class images.

Given the limited computation power, the fine-tuning process was done using a subset of 5300 images.
The two hyperparameters that are tuned are dropout rate (0.05,0.1,0.2) and learning rate (1e-4, 5e-4,
1e-3), the best set of parameters turns out to be dropout rate = 0.05 and learning rate = 5e-4. After the
model was fine-tuned, it was trained again using the large dataset of 21439 images.

Learning Rate Dropout Rate Accuracy Loss

0.0001 0.05 0.7826 0.4993

0.0005 0.05 0.7948 0.4943

0.001 0.05 0.7892 0.4937

0.0001 0.1 0.7771 0.5116

0.0005 0.1 0.7910 0.4944

0.001 0.1 0.7826 0.5066

0.0001 0.2 0.7715 0.5198

0.0005 0.2 0.7799 0.5028

0.001 0.2 0.7817 0.4897

Table 3. Results from hyperparameter tuning for MobileNetV2.

9
Detecting AI-Generated Fake Images

ResNet50 (Visualization displayed in Appendix 1.)

We used the ResNet50 model from the library Keras, pre-trained on the ImageNet dataset. We
removed the input and output layers and froze the rest of the weights since re-training the whole
network would be time-consuming. We then added an input layer of size (180, 180, 3). To the end of the
network, we added a Dense layer of size 512 and activation ReLU and an output Dense layer of size 1
and activation Sigmoid, since this is a binary classification problem. We also chose to use the Adam
optimizer and Binary Crossentropy as our loss function. We chose our batch size to be 300.

To get the best values for the hyperparameters learning rate and number of epochs, we performed
hyperparameter tuning using GridSearch. Since the training time was high for the full dataset, we
performed hyperparameter tuning on a subset of the data. For this process, we used the same ratio of
training and validation data with 4k training images and 1k validation images.

Learning Rate No. of Epochs Accuracy Precision Recall

0.01 10 0.85 0.93 0.74

0.001 10 0.86 0.91 0.8

0.0001 10 0.87 0.86 0.89

0.01 20 0.88 0.88 0.88

0.001 20 0.88 0.88 0.89

0.0001 20 0.88 0.87 0.9

0.01 50 0.88 0.87 0.9

Table 4. Results for hyperparameter tuning for ResNet50.

We limited hyperparameter tuning to the above hyperparameters and their values due to time

10
Detecting AI-Generated Fake Images

Figure 3. Accuracy vs. Epochs for ResNet50 with no. of epochs of 10 and learning rate smaller than 0.01.

constraints and large training times. We plan to include batch size, additional layers and input image
size to the process. Our initial aim was to pick the combination that gave us the best possible trade-off
between precision and recall (since we wanted to minimize misclassifying a real image as fake but at
the same time, identify as many fakes as possible) but there were a few challenges. We could not pick
epoch size greater than 10 due to increase in training time. We also noticed that the Accuracy vs.
Epoch graph for a learning rate smaller than 0.01 looked like Figure 3., where with more epochs the
accuracy wasn’t changing much, that is, the learning was too slow.

For these reasons, we chose the combination of learning rate of 0.01 and no. of epochs of 10. Had we
had access to more resources, we would have chosen no. of epochs to be 20.

For our final model, we used a training:validation ratio of the dataset to be 80:20, with 13k training
images, 3k validation images and 4k testing images.

Results
Classification Model

The training of the ResNet model to classify real and fake images is in progress. We will use the
accuracy metric, ROC curves, PR curves and a confusion matrix to evaluate the model and its
variations on a validation set. We will hold out a test set to get an estimate of the final generalization
performance after training on both the training and validation sets.

11
Detecting AI-Generated Fake Images

MobileNet

Evaluation

In total, 21439 images were used and separated into train sets (80%) and test sets (20%). The overall
accuracy is 80% with AUC being 0.88, precision being 0.74, recall being 0.91. Our aim was to get the
best possible trade-off between precision and recall since we wanted to minimize misclassifying a real
image as fake but at the same time, identify as many fakes as possible.

Precision Recall F1-score support

Real 0.88 0.65 0.75 2145

Fake 0.72 0.91 0.80 2143

Overall 0.8 0.78 0.78 4288

Table 5. Results for MobileNetV2.

Figure 4. Accuracy over the 10 epochs (left). Loss over the 10 epochs (right) of MobileNet V2 model.

Mapping

In this study, we consider four facial attributes: glasses, race (White, Asian, Black), emotion (positive,
negative, neutral), and gender (woman, man). Due to limitations in computing power, we randomly
selected 200 samples for each facial attribute and calculated the prediction accuracy of our model. The
results are summarized in Table 3.

Race
Asian 0.76
Black 0.82
White 0.80

12
Detecting AI-Generated Fake Images

Race
Asian 0.76
Hispanic 0.85
Gender
Man 0.83
Woman 0.87
Glasses
Glasses 0.74
No Glasses 0.78
Emotions
Positive 0.76
Neutral 0.78
Negative 0.75

Table 6. Accuracy for MobileNetV2 by facial category.

We see from the table above that our model does best for faces belonging to women under the
category gender, to race Asian under the category race, for faces without glasses and with neutral
expressions. After manually analyzing the fakes generated by the rest, we believe that DALL-E 2
performs worse when generating glasses with facial features like frowns or visible teeth in a smile.

ResNet50

Evaluation

In total, 13291 images were used for training, 3322 for validation and 4210 for testing with
hyperparameter values of 0.01 for learning rate and 10 for number of epochs. The overall accuracy is
89% with AUC of 0.96, precision of 0.86 and recall of 0.94 on the test set.

Accuracy AUC Precision Recall

ResNet50 0.89 0.96 0.86 0.94

Table 7. Results for ResNet50.

Real Fake

Real 1785.0 320

Fake 126 1979

Table 8. Confusion matrix for ResNet50.

13
Detecting AI-Generated Fake Images

Mapping

In this study, we consider four facial attributes: glasses, race (White, Asian, Black, Latino Hispanic),
emotion (positive, negative, neutral), and gender (woman, man). Due to limitations in computing power,
we randomly selected 200 samples for each facial attribute and calculated the prediction accuracy of
our model. The results are summarized in Table 6.

Race
Asian 0.835
Black 0.9
White 0.82
Hispanic 0.85
Gender
Man 0.85
Woman 0.84
Glasses
Glasses 0.80
No Glasses 0.84
Emotions
Positive 0.82
Neutral 0.86
Negative 0.825

Table 9. Accuracy for ResNet50 by facial category.

We see from the table above that our model does best for faces belonging to men under the category
gender, to race Black under the category race, for faces without glasses and with neutral expressions.
The results are similar for expressions and glasses for both models. After manually analyzing the fakes
generated by the rest, we believe that DALL-E 2 performs worse when generating glasses with facial
features like frowns or visible teeth in a smile.

Below we have included an example of a fake image generated by DALL-E 2 that both models could
not correctly classify. We believe that this is because it is a headshot with a neutral expression of the
face and a clean background.

14
Detecting AI-Generated Fake Images

Figure 5. Corresponding real image (left) and fake image generated by DALL-E 2 (right).

Analysis of ResNet50

To better understand the working of the ResNet50 model and analyze what parts of the image the
model uses to differentiate between real and fake images, we studied the outputs of the convolutional
layers, that is the feature maps of the model.

Figure 6 is one of the images we used to do this analysis. We chose this particular image for the poorly
reconstructed reflection in the glasses by DALL-E 2.

We then ran the image forward through our model and collected the feature maps at each layer. Figure
7 shows the plotted feature maps. We can see in the first few layer outputs that the model is initially
learning the features of a face, the shape, the eyes, nose, mouth, forehead, chin etc. (due to the
ResNet being trained on a larger, more general ImageNet dataset) but the learning in the last layers of
the model are too small to visualize and interpret.

For this reason, we constructed a heatmap using the output of the resnet’s Global Average Pooling
layer in Figure 8. We see that for this image the model does indeed focus on the poorly regenerated
reflection on the glasses and the sides of the face.

15
Detecting AI-Generated Fake Images

Figure 6. Corresponding real image (right) and fake image generated by DALL-E 2 (left).

Figure 7. Feature maps generated by ResNet50.

16
Detecting AI-Generated Fake Images

Figure 8. Heatmap generated by the Global Average Pooling layer of ResNet50.

17
Detecting AI-Generated Fake Images

Visualization
After we train the ResNet50, we deploy the pre-trained Keras model on HuggingFace, and use it for
visualization. Here, we cluster and plot the feature maps and their corresponding images from the final
layer of ResNet to be able to visualize the model’s predictions.

We retrieve a specified number of real and fake images, preparing them for processing. The images are
resized to a uniform size, and their pixel values are normalized to a range between 0 and 1.

To extract features from the images, we take the feature map generated by the final layer of the
pre-trained Keras model (ResNet50). The extracted features are then utilized to create the
visualization. The images are organized in a grid-like structure, with each image accompanied by a
label indication whether it is real or fake. The labels are color-coded, with red for fake and green for real
ones.

We see from the below plot that similar images (faces with similar attributes are grouped together) as
are the predicted fake or real images. From this we can conclude that the ResNet model learns the
attributes of the faces and whether the image is real or fake and this information is encoded in the
feature map of the final model layer.

Figure 9. Similar characteristics placed together.

18
Detecting AI-Generated Fake Images

Figure 10. Visualization result based on ResNet50 Model on higer demensional.

19
Detecting AI-Generated Fake Images

Dimensionality Reduction
We also tried the images in a lower-dimensional space (2D) obtained by UMAP, which helps in
identifying patterns and relations between the images. UMAP is an unsupervised dimensionality
reduction technique that preserves the global structure of the data while also revealing local pattern. In
the Figure 11, each image is placed based on its UMAP coordinates, and the border color represents
the cluster label from HDBSCAN (with separate color maps for real and fake images). UMAP is applied
to the extracted features to reduce their dimensionality to 2D, making it possible to visualize the images
in a 2D plot.
The plot shows how images are related to each other based on their extracted features, and the
color-coded borders help you understand which cluster each image belongs to and whether the image
is real or fake. The generated plot makes it easier to analyze the distribution of images and identify
clusters of similar images.

Figure 11. Visualization result based on ResNet50 Model on lower demensional using UMAP.

20
Detecting AI-Generated Fake Images

Deployment
After training the MobileNet model on a dataset of 21439 images stored in AWS S3, the trained models
was saved in the TensorFlow format and pushed to HuggingFace's model repositories (Emmawang /
mobilenet_v2_fake_image_detection and poojakabber1997/ResNetDallE2Fakes).

To enable practical usage of the model in real-world scenarios, the model was deployed using
Streamlit, an open-source Python library for building web applications, and hosted on HuggingFace
Space (spaces/Emmawang/Fake_image_detection). This enables users to interact with the model
through a user-friendly web interface and receive predictions on new images uploaded to the
application.

Deploying the model in this way enables wider accessibility and usability, allowing users to utilize the
model's capabilities without requiring specialized knowledge or expertise in machine learning.
HuggingFace Space provides an accessible and scalable platform for hosting models and making them
available to a wider audience:
https://huggingface.co/spaces/Emmawang/Fake_image_detection.

After thorough research and experimentation, our team has selected ResNet as the optimal model for
our web application. The ResNet model has been saved on HuggingFace and in order to seamlessly
integrate the model into our webapp, we have leveraged Streamlit, a cutting-edge tool for building and
deploying data-driven apps. Our webapp not only utilizes the ResNet model for AI picture detection,
providing a prediction and the corresponding confidence score, but also includes a DALL-E 2 AI picture
generation function, allowing users to generate their own unique images.

The webapp can be accessed at https://fake-image-generator-detector.streamlit.app/. Please note that

in order to generate AI pictures, users will need to apply for their own OpenAI API keys. We are excited
to offer this innovative tool to users and believe it will have a positive impact on a wide range of
industries, from marketing to entertainment.

Conclusion
In conclusion, our project aims to address the challenges and ethical implications of using AI-generated
facial images by developing a machine learning model that can distinguish between real human faces
and images generated by DALL-E 2. Through the analysis of a diverse dataset that includes real
human images from the FFHQ dataset and fake images generated by Dall-E 2, we hope to gain
valuable insights into the differences between AI-generated and real human faces. This research has
the potential to contribute to the development of responsible and ethical practices for the use of
AI-generated images in various fields, such as criminal detection and prevention.

The studies conducted by Sabithaa et al. (2020), Zhang et al. (2020), and He et al. (2016) provide
valuable insights into the development of a fake image detection model using convolutional neural

21
Detecting AI-Generated Fake Images

networks (CNNs). He et al. (2016) presented the use of a pre-trained ResNet model to address the
issue of vanishing gradients in deep neural networks, which can be fine-tuned and visualized in our
study to gain insights into the learned features and prediction process.

Limitation

1. Limited classification accuracy: Due to time constraints, we could not achieve an ideal
classification accuracy for the original data. With an accuracy of around 80% for each class
(glass, race, facial expression, gender), we had to train the model without optimizing it further.
To improve accuracy, a more extensive dataset with over 40k images for each class could be
analyzed, but this was not feasible within our time limitations.

2. Model comparison and selection: Although we compared different models, we did not have the
opportunity to explore the best possible model for our task. Combining MobileNetV2 and
ResNet might have yielded a more accurate and efficient model, but due to time constraints, this
possibility was not fully explored.

By addressing the challenges and ethical implications of using AI-generated images, this research
contributes to the development of responsible and ethical practices in various fields.

By drawing upon these studies, we can enhance the architecture and performance of our fake image
detection model. Our study aims not only to achieve high accuracy in detecting fake images, but also to
understand the underlying features and categories that significantly impact the model's performance,
such as race, gender, facial expressions, and wearing accessories. By incorporating insights from these
studies, our model can be further refined and optimized for real-world applications where the detection
of fake images generated by AI is crucial.

Roles
Our overall high-level roles include

1. Project management
2. Dataset acquisition
3. Building the model pipeline
4. Model training
5. Model experiments
6. Evaluation

We want to ensure that each member takes part in all the above roles. To break down, we first divided
the project and its deliverables into steps and tasks that can be assigned and distributed.

Tasks:
1. Pooja Kabber

22
Detecting AI-Generated Fake Images

a. Selecting the right model type and architecture

b. Experimenting with the model results
2. Scott Lai
a. Building the model pipeline
b. Experimenting with the model results
3. Song Young Oh
a. Training the model and model selection by hyperparameter optimization
b. Model evaluation
4. Emma Wang
a. Building / Finding and preparing the dataset
b. Model training/ fine tuning
c. Model evaluation

References
1. Sabithaa, R., Aruna, A., Karthik, S., & Shanthini, J. (2020). Enhanced model for fake image detection
(EMFID) using convolutional neural networks with histogram and wavelet based feature extractions.
Journal of Ambient Intelligence and Humanized Computing, 11(7), 3005-3018.

2. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).

3. Zeyang S., Zheng L., Ning Y., and Yang Z. (2022). DE-FAKE: Detection and Attribution of Fake
Images Generated by Text-to-Image Generation Models. arXiv preprint arXiv:2210.06998.

​4. Riccardo C., Davide C., Giada Z., Giovanni P., Koki N., and Luisa V. (2022). On the Detection of
Synthetic Images Generated by Diffusion Models. arXiv preprint arXiv:2211.00680v1.

5. Utkarsh O., Yuheng L., and Yong Jae L. (2023). Towards Universal Fake Image Detectors that
Generalize Across Generative Models. arXiv preprint arXiv:2302.10174v1.

6. Rezende, E.R.S.D.; Ruppert, G.C.S.; Carvalho, T. Detecting Computer Generated Images with Deep
Convolutional Neural Networks. In Proceedings of the 30th SIBGRAPI Conference on Graphics,
Patterns and Images, Niteroi, Brazil, 17–20 October 2017; pp. 71–78.

7. Ye Y., Weitong H., Wei Z., Ting W., and Yun-Qing S. (2018). Distinguishing Computer-generated
Graphics from Natural Images Based on Sensor Pattern Noise and Deep Learning. arXiv preprint
arXiv:1803.09403.

8. Bekhet S. & Alahmer H. A Robust Deep Learning Approach for Glass Detection in Non-standard
Facial Images. IET Biom. 2021;10:74-86.

9. Falko M., Christian R., and Marc S. Exploiting visual artifacts to expose deepfakes and face
manipulations. In 2019 IEEE Winter Applications of Computer Vision Workshops (WACVW), pages
83–92. IEEE, 2019.

23
Detecting AI-Generated Fake Images

10. Andreas R., Davide C., Luisa V., Christian R., Justus T., and Matthias N. Faceforensics++: Learning
to detect manipulated facial images. In Proceedings of the IEEE International Conference on Computer
Vision, pages 1–11, 2019.

11. Lingzhi L., Jianmin B., Ting Z., Hao Y., Dong C., Fang W., and Baining G. Face x-ray for more
general face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 5001–5010, 2020.

12. Yuyang Q., Guojun Y., Lu S., Zixuan C., and Jing S. Thinking in frequency: Face forgery detection
by mining frequency-aware clues. In the European Conference on Computer Vision, pages 86–103.
Springer, 2020.

13. Xi W., Zhen X., YuTao G., and Yu X. Sstnet: Detecting manipulated faces through spatial,
steganalysis and temporal features. In ICASSP 2020-2020 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pages 2952–2956. IEEE, 2020.

14. Xin Y., Yuezun L., and Siwei L. Exposing deep fakes using inconsistent head poses. In IEEE
International Confer- ence on Acoustics, Speech and Signal Processing, ICASSP 2019, Brighton,
United Kingdom, May 12-17, 2019, pages 8261–8265. IEEE, 2019.

15. Mingxing T. and Quoc V. Le. EfficientNet: Rethinking model scaling for convolutional neural
networks. In Ka- malika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th
International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California,
USA, volume 97 of Proceedings of Machine Learning Research, pages 6105–6114. PMLR, 2019.

16. Francois Chollet. Xception: Deep Learning with Depthwise Separable Convolutions. In IEEE
Conference on Computer Vision and Pattern Recognition, 2017.

17. Trpakov, D. (2022). Vit Face Expression Model. Hugging Face.

https://huggingface.co/trpakov/vit-face-expression

18. Siddharth Mandgi (2021). Glasses Detection - OpenCV, DLIB & Edge Detection. MLearning.ai.
https://medium.com/mlearning-ai/glasses-detection-opencv-dlib-bf4cd50856da

19. O. M. Parkhi, A. Vedaldi, A. Zisserman: Deep Face Recognition. British Machine Vision Conference,
2015

24
Detecting AI-Generated Fake Images

Appendix

Figure 10. Architecture of MobileNetV2.

25
Detecting AI-Generated Fake Images

Figure 11. Architecture of ResNet50

26