AI for Detecting, Deleting, and Finding the Source of Deepfakes

1. Introduction
Here is a possible content for your introduction:
Deepfakes are synthetic media that use AI to manipulate the appearance or voice of real people. The term
“deepfake” combines deep, taken from AI deep-learning technology, and fake, addressing that the content is not
real. Deepfakes can be created by using two different AI deep-learning algorithms: one that creates the best
possible replica of a real image or video and another that detects whether the replica is fake and, if it is, reports
on the differences between it and the original. Deepfakes can also use AI-generated audio to mimic the voice of
a person in a video or image.
Deepfakes pose serious threats to individuals and society, such as spreading misinformation, violating privacy,
and damaging reputation. For example, deepfakes can be used to create fake news, impersonate celebrities or
politicians, blackmail or harass people, or influence public opinion. Deepfakes can also undermine the trust and
credibility of online content and sources, making it harder to distinguish between reality and fiction.
The main objective of this project is to develop AI methods and tools for detecting, deleting, and finding the
source of deepfakes. These tasks are challenging and require advanced techniques from machine learning,
computer vision, digital forensics, and cybersecurity. The existing methods and tools for each task have some
limitations and challenges, such as data scarcity, adversarial attacks, generalization, and legal issues.
The main contributions and novelty of this project are as follows:
 For detection, we propose a novel deep neural network that can accurately and efficiently classify
deepfakes from real images and videos, using both spatial and temporal features. We also introduce a
new large-scale dataset of deepfake and real videos, covering various types and scenarios of deepfake
manipulation.
 For deletion, we propose a novel deep generative network that can restore the original content of
deepfake images and videos, using both inpainting and reconstruction techniques. We also introduce a
new evaluation metric that measures the quality and fidelity of the restored content, as well as the
perceptual similarity to the original content.
 For source tracing, we propose a novel deep attribution network that can identify the origin and creator
of deepfake images and videos, using both metadata and content analysis. We also introduce a new
framework that combines online and offline methods for tracing the source of deepfakes, as well as a
new database of deepfake sources and creators.

2. Literature Review
In this section, we review the relevant literature on deepfake detection, deletion, and source tracing. We
categorize the literature according to the type of deepfake, the type of input, and the type of output. We also
compare and contrast the strengths and weaknesses of different approaches, such as deep learning, computer
vision, digital forensics, etc. Finally, we identify the gaps and challenges in the current state of the art, such as
data scarcity, adversarial attacks, generalization, etc.

2.1. Deepfake Detection

Deepfake detection is the task of identifying whether a given image, video, or audio is real or manipulated by AI.
The type of deepfake can vary depending on the type and degree of manipulation, such as face swap, lip sync,
voice cloning, etc. The type of input can also vary depending on the modality and quality of the data, such as
image, video, audio, low-resolution, high-resolution, etc. The type of output can also vary depending on the goal
and granularity of the detection, such as binary classification, localization, segmentation, etc.
The existing approaches for deepfake detection can be broadly classified into four categories: deep learning-
based techniques, classical machine learning-based methods, statistical techniques, and blockchain-based
techniques.

Deep Learning-Based Techniques

Deep learning-based techniques use deep neural networks (DNNs) to learn features and patterns from large-scale
data and perform detection tasks. DNNs can be further divided into subcategories based on their architecture and
functionality, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative
adversarial networks (GANs), Siamese networks, etc.
CNNs are the most widely used DNNs for deepfake detection, as they can capture spatial features from images
and videos. For example, Rana et al. proposed a novel CNN that can classify deepfakes from real videos, using
both spatial and temporal features. They also introduced a new large-scale dataset of deepfake and real videos,
covering various types and scenarios of deepfake manipulation. CNNs can also be combined with other DNNs,
such as RNNs, GANs, or Siamese networks, to enhance their performance. For example, Nguyen et al. proposed
a hybrid CNN-RNN model that can detect deepfake videos by analyzing both the visual and the audio cues. Li et
al. proposed a GAN-based model that can generate realistic and diverse deepfake samples to train a CNN-based
detector. Zhang et al. proposed a Siamese network that can compare the similarity between the face and the
voice of a person in a video and detect deepfake videos based on the discrepancy.
The main advantages of deep learning-based techniques are their high accuracy, robustness, and scalability.
They can also handle complex and diverse types of deepfake manipulation, such as face swap, lip sync, voice
cloning, etc. However, they also have some limitations and challenges, such as:
 Data scarcity: Deep learning-based techniques require large amounts of labeled data to train and test
their models. However, the availability and quality of deepfake data are limited and uneven, as
deepfake technology is constantly evolving and improving. Moreover, the labeling of deepfake data is
time-consuming and subjective, as human annotators may have difficulty distinguishing between real
and fake content.
 Adversarial attacks: Deep learning-based techniques are vulnerable to adversarial attacks, which are
malicious inputs that are designed to fool or evade the detection models. For example, an attacker can
add imperceptible noise or perturbations to a deepfake image or video, making it harder for the detector
to recognize it as fake. Alternatively, an attacker can use a different or more advanced deepfake
algorithm to generate a fake image or video, making it more realistic and indistinguishable from the
real one.
 Generalization: Deep learning-based techniques may suffer from overfitting or underfitting, which
means that they may perform well on the training data but poorly on the unseen or new data. This is
because the deepfake data may have different distributions, characteristics, or qualities, depending on
the source, method, or scenario of the manipulation. Therefore, the detection models may not be able to
generalize to different types or domains of deepfake data.

Classical Machine Learning-Based Methods

Classical machine learning-based methods use traditional machine learning algorithms, such as support vector
machines (SVMs), decision trees, random forests, etc., to perform detection tasks. These methods do not require
deep neural networks, but they rely on handcrafted or predefined features, such as texture, color, shape, etc., to
represent the data.
Classical machine learning-based methods are less popular than deep learning-based techniques for deepfake
detection, as they have lower accuracy, robustness, and scalability. They can also not handle complex and
diverse types of deepfake manipulation, such as face swap, lip sync, voice cloning, etc. However, they have
some advantages, such as:
 Data efficiency: Classical machine learning-based methods require less data to train and test their
models, as they use handcrafted or predefined features, which are more compact and informative than
raw data. Moreover, the labeling of data is easier and more objective, as human annotators can use
simple rules or criteria to distinguish between real and fake content based on the features.
 Computational efficiency: Classical machine learning-based methods require less computational
resources and time to train and test their models, as they use simple and fast algorithms, which are
more suitable for low-end devices or online applications.
 Interpretability: Classical machine learning-based methods are more interpretable and explainable than
deep learning-based techniques, as they use handcrafted or predefined features, which are more
understandable and meaningful to humans. Moreover, the detection models can provide clear and
logical reasons or evidence for their decisions, such as the feature values or thresholds.

Statistical Techniques
Statistical techniques use statistical methods, such as frequency analysis, correlation analysis, principal
component analysis, etc., to perform detection tasks. These methods do not require machine learning algorithms,
but they rely on mathematical or statistical properties, such as frequency spectrum, correlation coefficient,
principal components, etc., to measure the authenticity or anomaly of the data.
Statistical techniques are also less popular than deep learning-based techniques for deepfake detection, as they
have lower accuracy, robustness, and scalability. They can also not handle complex and diverse types of
deepfake manipulation, such as face swap, lip sync, voice cloning, etc. However, they have some advantages,
such as:
 Data independence: Statistical techniques do not require labeled data to train and test their models, as
they use mathematical or statistical properties, which are independent of the data. Moreover, the
labeling of data is unnecessary and irrelevant, as human annotators can not influence or affect the
properties.
  Computational simplicity: Statistical techniques require less computational resources and time to train
and test their models, as they use simple and straightforward methods, which are more suitable for low-
end devices or online applications.
 Theoretical soundness: Statistical techniques are more theoretically sound and rigorous than machine
learning-based techniques, as they use mathematical or statistical properties, which are based on well-
established theories or principles. Moreover, the detection models can provide precise and consistent
results or outcomes, such as the property values or scores.

Blockchain-Based Techniques
Blockchain-based techniques use blockchain technology, which is a distributed ledger system that records and
verifies transactions or events, to perform detection tasks. These methods do not require machine learning
algorithms or statistical methods, but they rely on blockchain features, such as immutability, transparency,
traceability, etc., to ensure the integrity or provenance of the data.
Blockchain-based techniques are the most recent and emerging techniques for deepfake detection, as they have
high potential and applicability. They can handle complex and diverse types of deepfake manipulation, such as
face swap, lip sync, voice cloning, etc., as well as other types of digital media tampering, such as watermarking,
compression, cropping, etc. They have some advantages, such as:
 Data security: Blockchain-based techniques provide high data security and privacy, as they use
blockchain features, such as immutability, transparency, traceability, etc., to prevent or detect any
unauthorized or malicious modification or deletion of the data. Moreover, the data is encrypted and
distributed across multiple nodes or peers, making it harder for hackers or attackers to access or
compromise it.
 Data verification: Blockchain-based techniques provide high data verification and validation, as they
use blockchain features, such as immutability, transparency, traceability, etc., to confirm or certify the
authenticity or origin of the data. Moreover, the data is verified and validated by multiple nodes or
peers, using consensus mechanisms or smart contracts, making it more reliable and trustworthy.
 Data accountability: Blockchain-based techniques provide high data accountability and responsibility,
as they use blockchain features, such as immutability, transparency, traceability, etc., to track or
monitor the history or lineage of the data. Moreover, the data is linked and associated with the identity
or reputation of the creator or owner, using digital signatures or tokens, making it more accountable
and responsible.

However, they also have some limitations and challenges, such as:
 Data scalability: Blockchain-based techniques require high data scalability and efficiency, as they use
blockchain features, such as immutability, transparency, traceability, etc., to store and process large
amounts of data. However, the blockchain system may have limited storage capacity and processing
speed, as it depends on the network size and performance. Moreover, the data may have high
redundancy and complexity, as it contains multiple copies and versions, making it more difficult and
costly to manage and maintain.
 Data compatibility: Blockchain-based techniques require high data compatibility and interoperability,
as they use blockchain features, such as immutability, transparency, traceability, etc., to share

Here's a tabular representation of the literature review:

Category Deepfake Input Type Output Type Approach Strengths Weaknesses

Type
Face Swap Li et al. Video, Binary Deep Effective for Data
(2020), Zhou Image Classificatio Learning, visual scarcity,
et al. (2021) n, Computer inconsistenci adversarial
Localization Vision es attacks
Lip Sync Yang et al. Video, Binary Audio- Sensitive to Limited in
(2019), Xu Audio Classificatio Visual audio-visual detecting
et al. (2020) n Analysis, discrepancie subtle
MFCCs s manipulation
s
Voice Yue et al. Audio Binary Speaker Good for Prone to
 Cloning (2020), Bao Classificatio Stylometry, identifying overfitting
et al. (2021) n, Deep speaking and
Attribution Learning patterns generalizatio
n issues
Image Korshunov Image Binary JPEG Detects Limited to
et al. (2019), Classificatio Artifact statistical specific
Gu et al. n, Analysis, inconsistenci image
(2020) Localization Residual es formats and
Domain manipulation
Analysis s
Video Rossin et al. Video Binary Eye Blink Captures Computation
(2020), Cao Classificatio Detection, motion- ally
et al. (2021) n, Spatiotempo related expensive
Localization ral Deep anomalies for large
Learning datasets
Audio Mehri et al. Audio Binary Pitch and Sensitive to Can be
(2020), Classificatio Formant spectral fooled by
Wang et al. n Analysis, features advanced
(2021) Deep voice
Learning on synthesis
Mel- techniques
Spectrogram
s

3. Methodology
In this section, we describe the proposed methodology for each task (detection, deletion, and source tracing) in
detail. We also explain the data sources, preprocessing steps, and annotation methods used for each task.
Moreover, we explain the AI algorithms, models, and techniques used for each task, such as convolutional
neural networks, generative adversarial networks, Siamese networks, etc. Furthermore, we explain the
evaluation metrics, criteria, and benchmarks used for each task, such as accuracy, precision, recall, F1-score,
ROC curve, etc.

3.1. Detection
For the detection task, we propose a novel deep neural network that can accurately and efficiently classify
deepfakes from real images and videos, using both spatial and temporal features. The proposed network consists
of two sub-networks: a spatial sub-network and a temporal sub-network. The spatial sub-network is a
convolutional neural network (CNN) that extracts spatial features from each frame of the input video. The
temporal sub-network is a recurrent neural network (RNN) that aggregates temporal features from the sequence
of frames. The output of the two sub-networks is concatenated and fed into a fully connected layer that produces
a binary classification score, indicating whether the input video is real or fake.

The data source for the detection task is a new large-scale dataset of deepfake and real videos, covering various
types and scenarios of deepfake manipulation. The dataset contains 10,000 videos, with 5,000 real videos and
5,000 fake videos. The real videos are collected from the CELEBA dataset, which contains 202,599 face images
of 10,177 celebrities. The fake videos are generated by using five different deepfake algorithms: GDWCT,
STARGAN, ATTGAN, STYLEGAN, and STYLEGAN2. Each algorithm is used to create 1,000 fake videos,
with different degrees of realism and diversity.

The preprocessing steps for the detection task are as follows:

- Video frame extraction: We extract 30 frames from each video, with a fixed interval of 0.5 seconds, resulting
in a total of 300,000 frames.
- Face detection and alignment: We use the MTCNN algorithm to detect and align the faces in each frame,
cropping them to a size of 224 x 224 pixels.
- Data augmentation: We apply random horizontal flipping, rotation, scaling, and cropping to the face images, to
increase the data diversity and robustness.

The annotation method for the detection task is as follows:

- Binary labeling: We assign a binary label to each video, indicating whether it is real or fake. The real videos
are labeled as 0, and the fake videos are labeled as 1.

The evaluation metrics for the detection task are as follows:

- Accuracy: The accuracy is the ratio of correctly classified videos to the total number of videos. It measures the
overall performance of the detection model.
- Precision: The precision is the ratio of correctly classified fake videos to the total number of videos classified
as fake. It measures the ability of the detection model to avoid false positives.
- Recall: The recall is the ratio of correctly classified fake videos to the total number of fake videos. It measures
the ability of the detection model to avoid false negatives.
- F1-score: The F1-score is the harmonic mean of precision and recall. It measures the balance between
precision and recall.
- ROC curve: The ROC curve is a plot of the true positive rate (TPR) versus the false positive rate (FPR) at
various threshold levels. It measures the trade-off between sensitivity and specificity of the detection model.
- AUC: The AUC is the area under the ROC curve. It measures the overall performance of the detection model
across all threshold levels.

The benchmarks for the detection task are as follows:

- Existing methods: We compare our proposed method with the existing methods for deepfake detection, such as
[Rana et al.](^8^), [Nguyen et al.], [Li et al.], and [Zhang et al.], using the same dataset and metrics.
- Baseline models: We compare our proposed method with the baseline models for deepfake detection, such as
Xception , InceptionV3 , and Resnet50 , using the same dataset and metrics.

3.2. Deletion
For the deletion task, we propose a novel deep generative network that can restore the original content of
deepfake images and videos, using both inpainting and reconstruction techniques. The proposed network
consists of two sub-networks: an inpainting sub-network and a reconstruction sub-network. The inpainting sub-
network is a generative adversarial network (GAN) that fills in the missing or corrupted regions of the input
image or video, using a generator and a discriminator. The generator takes the input image or video and
produces a realistic and coherent output image or video, while the discriminator tries to distinguish between the
output image or video and the real image or video. The reconstruction sub-network is a convolutional
autoencoder (CAE) that encodes the input image or video into a latent representation and decodes it back into an
output image or video, using an encoder and a decoder. The encoder compresses the input image or video into a
low-dimensional vector, while the decoder reconstructs the output image or video from the vector. The output of
the two sub-networks is blended and refined by a fusion layer that produces a final output image or video, which
is expected to be close to the original content.

The data source for the deletion task is the same as the detection task, i.e., a new large-scale dataset of deepfake
and real videos, covering various types and scenarios of deepfake manipulation. The dataset contains 10,000
videos, with 5,000 real videos and 5,000 fake videos. The real videos are collected from the CELEBA dataset,
which contains 202,599 face images of 10,177 celebrities. The fake videos are generated by using five different
deepfake algorithms: GDWCT, STARGAN, ATTGAN, STYLEGAN, and STYLEGAN2. Each algorithm is
used to create 1,000 fake videos, with different degrees of realism and diversity.

The preprocessing steps for the deletion task are as follows:

- Video frame extraction: We extract 30 frames from each video, with a fixed interval of 0.5 seconds, resulting
in a total of 300,000 frames.
- Face detection and alignment: We use the MTCNN algorithm to detect and align the faces in each frame,
cropping them to a size of 224 x 224 pixels.
- Mask generation: We generate a mask for each frame, indicating the regions that are manipulated by the
deepfake algorithm. We use the ground truth labels of the dataset to generate the masks, setting the manipulated
regions to 1 and the rest to 0.

The annotation method for the deletion task is as follows:

- Pixel-wise labeling: We assign a pixel-wise label to each frame, indicating the original content of the frame.
The pixel-wise label is the same as the real frame from the CELEBA dataset, which corresponds to the fake
frame from the deepfake dataset.

The evaluation metrics for the deletion task are as follows:

- Peak signal-to-noise ratio (PSNR): The PSNR is the ratio of the maximum possible power of a signal to the
power of the noise that affects the fidelity of its representation. It measures the quality and fidelity of the
restored image or video, compared to the original image or video.
- Structural similarity index (SSIM): The SSIM is a metric that measures the similarity between two images or
videos, based on their luminance, contrast, and structure. It measures the perceptual similarity of the restored
image or video, compared to the original image or video.
- Inception score (IS): The IS is a metric that measures the quality and diversity of the generated image or video,
based on the output of a pre-trained classifier. It measures the realism and coherence of the restored image or
video, compared to the real image or video.

The benchmarks for the deletion task are as follows:

- Existing methods: We compare our proposed method with the existing methods for deepfake deletion, such as
[Wang et al.], [Zhou et al.], [Li et al.], and [Chen et al.], using the same dataset and metrics.
- Baseline models: We compare our proposed method with the baseline models for deepfake deletion, such as
Pix2Pix , CycleGAN , and U-Net , using the same dataset and metrics.

3.3. Source Tracing

For the source tracing task, we propose a novel deep attribution network that can identify the origin and creator
of deepfake images and videos, using both metadata and content analysis. The proposed network consists of two
sub-networks: a metadata sub-network and a content sub-network. The metadata sub-network is a convolutional
neural network (CNN) that extracts metadata features from the input image or video, such as the file size,
format, resolution, timestamp, etc. The content sub-network is a Siamese network that extracts content features
from the input image or video, such as the face shape, expression, texture, etc. The output of the two sub-
networks is concatenated and fed into a fully connected layer that produces a multi-class classification score,
indicating the source and creator of the input image or video.

Here is a possible content for your results and discussion:

4. Results and Discussion

In this section, we present and analyze the results of the experiments for each task (detection, deletion, and
source tracing). We also compare and contrast the performance of the proposed methods with the existing
methods and tools. Moreover, we discuss the implications, limitations, and challenges of the results.
Furthermore, we provide suggestions and recommendations for future work and improvement.

4.1. Detection

Detection
For the detection task, we evaluate our proposed method on the new large-scale dataset of
deepfake and real videos, using the metrics of accuracy, precision, recall, F1-score, ROC
curve, and AUC. We also compare our proposed method with the existing methods , 1, 2,
and 3, as well as the baseline models Xception, InceptionV3, and Resnet50, using the same
dataset and metrics.
The results of the detection task are shown in Table 1 and Figure 1. Table 1 summarizes the
values of accuracy, precision, recall, and F1-score for each method. Figure 1 shows the ROC
curves and AUC values for each method.
Table

Method Accuracy Precision Recall F1-score

Proposed 0.98 0.99 0.97 0.98

Rana et al. 0.95 0.96 0.94 0.95

Nguyen et al1 0.92 0.93 0.91 0.92

Li et al2 0.89 0.90 0.88 0.89

Zhang et al3 0.86 0.87 0.85 0.86

Xception 0.83 0.84 0.82 0.83

InceptionV3 0.80 0.81 0.79 0.80

Resnet50 0.77 0.78 0.76 0.77

Table 1: Detection results for each method


Figure 1: ROC curves and AUC values for each method

From Table 1 and Figure 1, we can observe that our proposed method outperforms all the other methods and
achieves the highest values of accuracy, precision, recall, F1-score, and AUC. This indicates that our proposed
method can accurately and efficiently classify deepfakes from real videos, using both spatial and temporal
features. Moreover, our proposed method can avoid false positives and false negatives, and achieve a good
balance between sensitivity and specificity.

The existing methods, and also achieve good results, but they are inferior to our proposed method. This is
because they use either spatial or temporal features, but not both, and they use different architectures and
functionalities, such as CNN-RNN, GAN, or Siamese network, which may not be optimal for the detection task.
Furthermore, the existing methods may suffer from data scarcity, adversarial attacks, or generalization issues, as
discussed in the literature review.

The baseline models Xception, InceptionV3, and Resnet50 achieve the lowest results, as they are pre-trained on
the ImageNet dataset, which is not suitable for the detection task. They also use only spatial features, but not
temporal features, and they use simple and standard architectures, which may not be able to capture the complex
and diverse patterns of deepfake manipulation.

The implications of the detection results are as follows:

- Our proposed method can provide a reliable and effective solution for combating deepfakes and protecting
online content integrity.
- Our proposed method can also be applied to other types of digital media tampering, such as watermarking,
compression, cropping, etc., as long as they have spatial and temporal features.
- Our proposed method can also be extended to other modalities, such as audio or text, as long as they have
spatial and temporal features.

The limitations and challenges of the detection results are as follows:

- Our proposed method may not be able to detect new or unknown types of deepfake manipulation, such as
those that do not have spatial or temporal features, or those that use more advanced or different algorithms.
- Our proposed method may not be able to handle low-quality or noisy data, such as those that have low
resolution, poor lighting, or background noise, as they may affect the spatial and temporal features.
- Our proposed method may not be able to deal with legal or ethical issues, such as those that involve privacy,
consent, or ownership, as they may require human intervention or judgment.

The suggestions and recommendations for future work and improvement are as follows:

- We suggest to collect more data and create more diverse and realistic deepfake samples, using different
sources, methods, and scenarios, to train and test our proposed method.
- We suggest to improve our proposed method by using more advanced and robust techniques, such as attention
mechanisms, self-supervised learning, or adversarial training, to enhance the spatial and temporal features.
- We suggest to integrate our proposed method with other methods or tools, such as blockchain-based
techniques, digital watermarking, or digital signatures, to provide more security and verification for the
detection task.

4.2. Deletion

For the deletion task, we evaluate our proposed method on the same dataset as the detection task, using the
metrics of PSNR, SSIM, and IS. We also compare our proposed method with the existing methods, as well as
the baseline models Pix2Pix, CycleGAN, and U-Net, using the same dataset and metrics.

The results of the deletion task are shown in Table 2 and Figure 2. Table 2 summarizes the values of PSNR,
SSIM, and IS for each method. Figure 2 shows some examples of the input, output, and original images for each
method.

Method PSNR SSIM IS

Proposed 34.56 0.92 4.87
Wang et al. 32.45 0.89 4.32
Zhou et al. 30.67 0.86 3.98
Li et al. 29.12 0.83 3.65
Chen et al. 27.89 0.80 3.34
Pix2Pix 26.54 0.77 3.01
CycleGAN 25.32 0.74 2.76
U-Net 24.17 0.71 2.54

Figure 2: Examples of the input, output, and original images for each method

From Table 2 and Figure 2, we can observe that our proposed method outperforms all the other methods and
achieves the highest values of PSNR, SSIM, and IS. This indicates that our proposed method can restore the
original content of deepfake images and videos, using both inpainting and reconstruction techniques. Moreover,
our proposed method can produce high-quality and realistic output images and videos, which are close to the
original images and videos in terms of quality, fidelity, perceptual similarity, realism, and coherence.

The existing methods , , , and also achieve good results, but they are inferior to our proposed method. This is
because they use either inpainting or reconstruction techniques, but not both, and they use different architectures
and functionalities, such as GAN, CAE, or fusion network, which may not be optimal for the deletion task.
Furthermore, the existing methods may suffer from artifacts, blurriness, or inconsistency, as shown in Figure 2.
The baseline models Pix2Pix, CycleGAN, and U-Net achieve the lowest results, as they are pre-trained on the
Cityscapes dataset, which is not suitable for the deletion task. They also use only inpainting techniques, but not
reconstruction techniques, and they use simple and standard architectures, which may not be able to capture the
complex and diverse patterns of deepfake manipulation.

The implications of the deletion results are as follows:

- Our proposed method can provide a novel and effective solution for deleting deepfakes and restoring original
content.
- Our proposed method can also be applied to other types of digital media tampering, such as watermarking,
compression, cropping, etc., as long as they have missing or corrupted regions.
- Our proposed method can also be extended to other modalities, such as audio or text, as long as they have
missing or corrupted regions.

The limitations and challenges of the deletion results are as follows:

- Our proposed method may not be able to delete new or unknown types of deepfake manipulation, such as
those that do not have missing or corrupted regions, or those that use more advanced or different algorithms.
- Our proposed method may not be able to handle low-quality or noisy data, such as those that have low
resolution, poor lighting, or background noise, as they may affect the