Available via license: CC BY 4.0
Content may be subject to copyright.
Information Technology and Control 2024/2/53
408
GAN-Generated Face
Detection Based on Multiple
Attention Mechanism and
Relational Embedding
ITC 2/53
Information Technology
and Control
Vol. 53 / No. 2 / 2024
pp. 408-428
DOI 10.5755/j01.itc.53.2.35590
GAN-Generated Face Detection Based on Multiple Attention
Mechanism and Relational Embedding
Received 2023/11/11 Accepted after revision 2024/03/02
HOW TO CITE: Ouyang, J., Ma, J., Chen, B. (2024). GAN-Generated Face Detection Based on
Multiple Attention Mechanism and Relational Embedding. Information Technology and Control,
53(2), 408-428. https://doi.org/10.5755/j01.itc.53.2.35590
Corresponding author: yangjunlin0732@163.com
Junlin Ouyang
School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, China;
Hunan Key Laboratory for Service computing and Novel Software Technology, Xiangtan, China;
Hunan Software Vocational and Technical University, Xiangtan, China; e-mail: yangjunlin0732@163.com
Jiayong Ma
School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, China;
Hunan Key Laboratory for Service computing and Novel Software Technology, Xiangtan, China
Beijing Chen
School of Computer and Software, Engineering Research Center of Digital Forensics, Ministry of Education,
Nanjing University of Information Science and Technology, Nanjing, China
The rapid development of the Generative Adversarial Network (GAN) makes generated face images more and
more visually indistinguishable, and the detection performance of previous methods will degrade seriously
when the testing samples are out-of-sample datasets or have been post-processed. To address the above prob-
lems, we propose a new relational embedding network based on “what to observe” and “where to attend” from a
relational perspective for the task of generated face detection. In addition, we designed two attention modules
to eectively utilize global and local features. Specifically, the dual-self attention module selectively enhances
the representation of local features through both image space and channel dimensions. The cross-correlation
attention module computes similarity between images to capture the global information of the output in the
image. We conducted extensive experiments to validate our method, and the proposed algorithm can eectively
extract the correlations between features and achieve satisfactory generalization and robustness in generating
409
Information Technology and Control 2024/2/53
face detection. In addition, we also explored the design of the model structure and the inspection performance
on more categories of generated images (not limited to faces). The results show that RENet also has good detec-
tion performance on datasets other than faces.
KEYWORDS: Image forensics; forgery detection; GAN-generated face detection; generative adversarial net-
works; relational networks.
1. Introduction
The Generative Adversarial Network (GAN) [18] has
been gradually applied to many fields since it was pro-
posed [34, 39, 57]. With its rapid development, the
generated images are becoming more and more real-
istic, as shown in Figure 1 for the face images gener-
ated by GAN. Such GAN-generated face images are
dicult to distinguish from human eyes and can be
easily generated by ordinary people. If they are used
for malicious purposes, it may adversely aect some
individuals’ reputations and even social security eth-
ics. Therefore, the detection of GAN-generated face
images has become increasingly necessary.
In recent years, some researchers [27, 50, 51] have
verified the authenticity of images by actively embed
watermarks to the images. Besides, many passive
forensic methods have been proposed. In [1, 7, 16,
19, 23, 45, 49, 59] they detect natural and generated
faces by exploring the dierences in the image for-
mation process. They combine traditional forensic
methods to generated face detection. However, when
confronted with fake face images where only local
areas are generated, searching for feature dierenc-
es directly on the entire generated face image may
lead to detection failure. Therefore, [4, 5, 8, 38] also
combine local information such as artifacts to assist
in detection. Although the methods described above
achieve relatively high detection accuracy, they suf-
fer from poor generalization and a lack of interpret-
ability [21]. [21, 22, 26, 64] try to make the results
interpretable by looking for inconsistencies in the
physiology-based methods. However, with the con-
tinuous innovation of GAN, the dierence between
generated images and natural images in the spatial
domain becomes increasingly dicult to detect [68].
As a consequence, [6, 9, 13, 14, 17, 37, 43, 68] turned
their attention to the frequency domain, which im-
proves the detection generalization performance by
fusing features from the spatial and frequency do-
mains. But their methods cannot adaptively capture
the most dis-criminative features.
To address the above challenges, we propose a method
that is inspired by relational networks [56] and com-
bines both dual-self attention (DSA) and cross-cor-
relation attention (CCA) to learn “what to observe”
and “where to attend” from an image relations per-
spective. We achieve this goal by utilizing relational
patterns within and between images through the Re-
lational Embedding Network (RENet).
The DSA module learns its own feature associations
for the purpose of enriching semantic in-formation.
The CCA module calculates the correlations between
images so that they can have global information. Our
contribution can be summarized as follows:
1 We propose a Relation Embedding Network (REN-
et) with a multi-attention mechanism to improve
the ability of detecting GAN-generated face.
2 We design dual-self attention and cross-correla-
tion attention to enhance the local spatial and
channel-wise correlated features within an image,
and to link global relationships between images,
respectively.
3 We conduct extensive experiments to verify the
proposed method has excellent generalization
ability and robustness to common post-processing
operations.
Figure 1
Some samples in the experimental datasets. From left to
right, the columns are the fake faces generated by ProGAN
[30], StyleGAN [31], StyleGAN2 [32], StarGAN [11],
BeGAN [2], LsGAN [46], WgGANGP [20], RelGAN [42]
respectively. The generated faces are indistinguishable by
human eyes
Information Technology and Control 2024/2/53
410
2. Related Work
Fake face detection. With the development of GAN
technology and fake face technology, researchers
have proposed many methods to detect fake faces. Hu
et al. [26] and Guo et al. [21] found that the highlight
of the two eyes in GAN-generated faces are inconsis-
tent. Guo et al. [22] found that the pupil shape of real
faces should be close to circular or elliptical, while
the pupils of generated faces show irregular and in-
consistent shapes. Although physiology-based incon-
sistency detection is interpretable, accuracy is great-
ly reduced if the inconsistencies are occluded or the
scene angle is biased. Nataraj et al. [49] extracted the
co-occurrence matrices on the RGB channels of the
image and input them into a neural network for clas-
sification. Barni et al. [1] reported a significant impact
of the correlation between color channels on detec-
tion eectiveness. Besides individual RGB channels,
they also calculated the co-occurrence matrices from
pairwise combinations of channels. The experimen-
tal results showed that multiple channels can further
improve the robustness of detection compared to us-
ing single color channel information. However, the
accuracy and robustness of detection only in the RGB
domain are far from satisfactory. Chen et al. [7] com-
bined dual color spaces and designed an improved
Xception network model to increase detection ro-
bustness. Guo et al. [23] adaptive convolution to pre-
dict manipulation traces in an image, and then max-
imized manipulation artifacts by updating weights
through backpropagation. Liu et al. [45] analyzed that
it was more robust to detect fake faces by texture, so
they proposed GramNet to capture long-range tex-
ture information to improve the robustness and gen-
eralization of the model. Despite the commendable
detection performance exhibited by previous meth-
ods, relying solely on the spatial domain still presents
limitations [14]. Moreover, the frequency domain of
images has been widely applied in various fields [10,
54, 69, 71, 72]. Frank et al. [14] found that there are
significant dierences in the DCT spectra of real and
fake images, and the DCT spectrum is more robust
for detecting image manipulation than the RGB spec-
trum. Liu et al. [43] argued that upsampling is a nec-
essary step in most face-forgery techniques, which
leads to significant changes in the frequency domain,
particularly the phase spectrum. Thus, they captured
upsampling artifacts in face-forgery by combining
spatial images with phase spectra, achieving a good
detection result. Luo et al. [67] discovered that noise
in face regions is continuously distributed in real im-
ages, while in manipulated images, it appears smooth-
er or sharper. Therefore, they employed the high-pass
filter SRM to extract high-frequency noise for detect-
ing face forgery. Le et al. [35] utilized frequency-do-
main knowledge distillation to retrieve the removed
high-frequency components in the student network
for enhancing the detection accuracy of low-resolu-
tion images. Although analyzing in other domains can
improve the accuracy and robustness of detection,
they focus on the global features of the image and
are typically dicult to detect subtle local tamper-
ing. Chai et al. [4] proposed the patchCNN network,
which truncates the entire network to focus on local
artifacts. Experiments have shown that local texture
information can enhance the model’s generalization
ability. Jia et al. [28] designed a dual-branch network
for predicting image-level and pixel-level fake labels
based on inter-image and intra-image inconsistency,
which is processed by stable wavelet decomposition.
Ju et al. [5] introduced the FPN modules into Xcep-
tion and reduced the number of convolutional lay-
ers in the network to detect locally generated face.
The model has good performance for detecting faces
with small generated regions. However, they overem-
phasized local features and ignored the relationship
between global and local features. In contrast to the
above, we integrate both global and local features,
making the framework more robust and demonstrat-
ing better generalization ability.
Attention model. The human visual nerve receives
more data than it can process, so it requires the hu-
man brain to weigh the inputs and focus only on the
necessary information [24]. For this reason, the
researchers used a similar concept in their experi-
ments. Vaswani et al. [58] first proposed a self-atten-
tion mechanism and applied it to machine transla-
tion, which reveals concerns about image structure
through similarities within a domain. Recent work
[40, 41, 48, 52, 62] has shown that the self-attention
mechanism can eectively capture contextual rela-
tionships and improve the intra-class compactness
of images. In addition to focusing on the connections
between images, the relationships between images
have been used to form a central part of various prob-
411
Information Technology and Control 2024/2/53
lems in computer vision. It calculates the relationship
between two images and applies to video action rec-
ognition, few-shot classification, semantic segmenta-
tion, medical image segmentation, style transfer, etc.
Recently, some GAN-generated detection algorithms
[5, 7] have adopted self-attention mechanisms to en-
hance semantic information, but they have not ac-
counted for connections between images. Inspired by
Wu et al. [63], we introduce not only self-attention in
each image but also cross-attention between images
to enhance the ability to classify relevant regions.
Relation Network. The relational network [56] is a
metric-based network structure with the core idea of
mapping images to a learnable embedding function to
extract features of interest and then distinguishing
dierent classes by measuring the similarity of fea-
tures between samples. Thus, attention can be used
to correct and strengthen the feature regions of inter-
est. In this paper, multiple attention mechanisms that
are employed post the embedding function to rectify
the network’s focus on relevant areas and boost the
expression of pertinent regions This approach pro-
motes the network’s generalization and robustness.
3. Approach
In this section, we will introduce the Relational Em-
bedding Network (RENet), which is used to solve
the problem of poor generalization and robustness
in GAN-generated face detection. In Figure 2, the
overall architecture consists of three modules: a basic
representation module, a feature augmentation mod-
ule, and a representation comparison module. The
network parameters of feature augmentation module,
and a representation comparison module are shown
in Table 2 The basic representation module and the
representation comparison module are similar to the
modules of the relational network. On this basis, we
add a feature enhancement module, which is com-
posed of dual-self attention (DSA) and cross-correla-
tion attention (CCA). We provide a brief overview of
the overall RENet in Section 3.1. Then, we introduce
the implementation details of DSA and CCA in Sec-
tion 3.2 and Sec. 3.3, respectively.
3.1. Architecture Overview
In this paper, we treat real images from the same da-
taset or fake images generated by same type of GAN
as one domain, which is similar to the setting of oth-
er related tasks [19, 58]. The training set for each
domain is denoted Dtrain, and the test set is denoted
Dtest. Both Dtrain and Dtest are divided into multiple ep-
isodes for training, each of them contains a query set
Q = (Iq, yq) and a support set S = {(IS
L, yS
L)}NK
L=1 with N cat-
egories and K images per category.
As shown in Figure 2, given a pair of support and query
set images {IS
L, Iq}, each of them has a size of C×H×W.
They pass through a shared embedding network and
generate corresponding features fs1, ..., fS
NK and fq.The
Figure 2
Overall architecture of RENet. We feed fS and fq from shared embedding network into the DSA module to obtain locally
enhanced features fs
Dual and fq
Dual. Following CCA, we can obtain FS and Fq, both of which contain global information. Finally,
they are concatenated along the channel and the category with the highest score determines the classification result
Information Technology and Control 2024/2/53
412
support set features fs1, ..., fS
NK from the same domain
will be ∈ denoted as fs via element-wise sum. The DSA
module first applies self-attention over them to gen-
erate self-attentive features {As
Dual, Aq
Dual} ∈ ℝC×H×W.
and multiplies them by their corresponding weights
before adding them to the input features to obtain fs-
Dual and fqDual, respectively. The resulting features are
then processed by the CCA module to generate {As
Cross,
Aq
Cross) ∈ ℝC×H×W, which operates similarly to the DSA
module. From there, output feature Fs and Fq are com-
puted for the classification score by representation
comparison, with Fs being calculated as follows:
Overall architecture of RENet. We feed and from shared embedding network into the DSA module to obtain
locally enhanced features and . Following CCA, we can obtain and , both of which contain global
information. Finally, they are concatenated along the channel and the category with the highest score determines the
classification result.
As shown in Figure 2, given a pair of support and
query set images
, each of them has a size of
C×H×W. They pass through a shared embedding net-
work and generate corresponding features
and .The support set features
from the same domain will be denoted
as via element-wise sum. The DSA module first
applies self-attention over them to generate self-atten-
tive features
Dual . and multi-
plies them by their corresponding weights before add-
ing them to the input features to obtain and
, respectively. The resulting features are then
processed by the CCA module to generate
Cross, which operates similarly to the
DSA module. From there, output feature and
are computed for the classification score by represen-
tation comparison, with being calculated as fol-
lows:
(1)
where and are learnable parameters for as-
signing weights, which gradually learn a weight from
0. The output features have their own local enhance-
ments as well as correlation information between im-
ages. is calculated in a similar way to .
3.2. Dual-Self Attention (DSA)
Self-Position Attention(SPA) Global texture fea-
tures help to detect GAN-generated faces and can im-
prove model detection ability by capturing long-range
texture information[15]. However, [16, 17] show that
generated face detection relies more on subtle details
in the hair or background when performing cross da-
taset detection, so focusing on local detailed features
can effectively enhance the generalization ability of
detection. In order to pay more attention to local de-
tailed features, we introduce the self-position atten-
tion module, which lets local features establish links
in contextual relationships, thereby enhancing the net-
work's expression of local features.
As shown in the upper part of Figure 3, given the fea-
ture , it is first fed into a 1x1 convolu-
tional layer to generate , where (
). We then reshape and into
and , where in the
SPA module. Next, is transposed to
, and the matrix multiplication is performed
between and to obtain their relationship, de-
noted by . After softmax layer, we will ob-
tain position attention maps . From the
spatial position， of and of , we respec-
tively get two spatial points {
}
,where ，. We fur-
ther denote the pointwise calculation of ) as
, and have the following equation:
(2)
where represents feature at position i
impact on position j. The more correlated the features
at the two positions are, the larger is. Then, after
multiplying with the matrix
, we sum
them to obtain the correlation of all features in column
i, :
. (3)
Finally, we aggregate the features of all points and re-
shape to get the final output i.e.:
(1)
where λ1 and θ1 are learnable parameters for assigning
weights, which gradually learn a weight from 0. The
output features have their own local enhancements as
well as correlation information between images. Fq is
calculated in a similar way to Fs.
3.2. Dual-Self Attention (DSA)
Self-Position Attention(SPA). Global texture fea-
tures help to detect GAN-generated faces and can im-
prove model detection ability by capturing long-range
texture information [15]. However, [16, 17] show
that generated face detection relies more on subtle
details in the hair or background when performing
cross dataset detection, so focusing on local detailed
features can eectively enhance the generalization
ability of detection. In order to pay more attention to
local detailed features, we introduce the self-position
attention module, which lets local features establish
links in contextual relationships, thereby enhancing
the network’s expression of local features.
As shown in the upper part of Figure 3, given the fea-
ture f ∈ ℝC×H×W, it is first fed into a 1x1 convolutional
layer to generate f1 ∈ ℝC'×H×W where (C'<C ). We then
reshape f1 and f into f1' ∈ ℝC'×N and f2'∈ ℝC×N, where
N = H × W in the SPA module. Next, f1' is transposed
to f1'T∈ ℝN×C', and the matrix multiplication is per-
formed between f1' and f1'T to obtain their relation-
ship, denoted by g(f1', f1'T ). After softmax layer, we
will obtain position attention maps S ∈ ℝN×N. From
the spatial position, i of f1' and j of f1'T, we respective-
ly get two spatial points {f1i
', f1j
'T} ℝC',where i ∈ {1, ...,
N}, j ∈ {1, ..., N} . We further denote the pointwise
calculation of g(f1', f1'T ) as gij(f1i
', f1j
'T ), and have the
following equation:
Figure 3
The details of DSA. Features are fed to the self-position attention (SPA, upper) and self-channel attention (SCA, below)
modules, which output Ep and Ec, respectively. They then fuse into features ADual with positional attention and channel
attention. DSA aims to make the network better understand “what to observe”
413
Information Technology and Control 2024/2/53
Overall architecture of RENet. We feed and from shared embedding network into the DSA module to obtain
locally enhanced features and . Following CCA, we can obtain and , both of which contain global
information. Finally, they are concatenated along the channel and the category with the highest score determines the
classification result.
As shown in Figure 2, given a pair of support and
query set images
, each of them has a size of
C×H×W. They pass through a shared embedding net-
work and generate corresponding features
and .The support set features
from the same domain will be denoted
as via element-wise sum. The DSA module first
applies self-attention over them to generate self-atten-
tive features
Dual . and multi-
plies them by their corresponding weights before add-
ing them to the input features to obtain and
, respectively. The resulting features are then
processed by the CCA module to generate
Cross, which operates similarly to the
DSA module. From there, output feature and
are computed for the classification score by represen-
tation comparison, with being calculated as fol-
lows:
(1)
where and are learnable parameters for as-
signing weights, which gradually learn a weight from
0. The output features have their own local enhance-
ments as well as correlation information between im-
ages. is calculated in a similar way to .
3.2. Dual-Self Attention (DSA)
Self-Position Attention(SPA) Global texture fea-
tures help to detect GAN-generated faces and can im-
prove model detection ability by capturing long-range
texture information[15]. However, [16, 17] show that
generated face detection relies more on subtle details
in the hair or background when performing cross da-
taset detection, so focusing on local detailed features
can effectively enhance the generalization ability of
detection. In order to pay more attention to local de-
tailed features, we introduce the self-position atten-
tion module, which lets local features establish links
in contextual relationships, thereby enhancing the net-
work's expression of local features.
As shown in the upper part of Figure 3, given the fea-
ture , it is first fed into a 1x1 convolu-
tional layer to generate , where (
). We then reshape and into
and , where in the
SPA module. Next, is transposed to
, and the matrix multiplication is performed
between and to obtain their relationship, de-
noted by . After softmax layer, we will ob-
tain position attention maps . From the
spatial position， of and of , we respec-
tively get two spatial points {
}
,where ，. We fur-
ther denote the pointwise calculation of ) as
, and have the following equation:
(2)
where represents feature at position i
impact on position j. The more correlated the features
at the two positions are, the larger is. Then, after
multiplying with the matrix
, we sum
them to obtain the correlation of all features in column
i, :
. (3)
Finally, we aggregate the features of all points and re-
shape to get the final output i.e.:
(2)
where sij ∈ ℝ1×1 represents feature at position i impact
on position j. The more correlated the features at the
two positions are, the larger sij is. Then, after multiply-
ing with the matrix f2i
' ∈ ℝ1×N, we sum them to obtain
the correlation of all features in column i, Ej ∈ ℝ1×N:
Overall architecture of RENet. We feed and from shared embedding network into the DSA module to obtain
locally enhanced features and . Following CCA, we can obtain and , both of which contain global
information. Finally, they are concatenated along the channel and the category with the highest score determines the
classification result.
As shown in Figure 2, given a pair of support and
query set images
, each of them has a size of
C×H×W. They pass through a shared embedding net-
work and generate corresponding features
and .The support set features
from the same domain will be denoted
as via element-wise sum. The DSA module first
applies self-attention over them to generate self-atten-
tive features
Dual . and multi-
plies them by their corresponding weights before add-
ing them to the input features to obtain and
, respectively. The resulting features are then
processed by the CCA module to generate
Cross, which operates similarly to the
DSA module. From there, output feature and
are computed for the classification score by represen-
tation comparison, with being calculated as fol-
lows:
(1)
where and are learnable parameters for as-
signing weights, which gradually learn a weight from
0. The output features have their own local enhance-
ments as well as correlation information between im-
ages. is calculated in a similar way to .
3.2. Dual-Self Attention (DSA)
Self-Position Attention(SPA) Global texture fea-
tures help to detect GAN-generated faces and can im-
prove model detection ability by capturing long-range
texture information[15]. However, [16, 17] show that
generated face detection relies more on subtle details
in the hair or background when performing cross da-
taset detection, so focusing on local detailed features
can effectively enhance the generalization ability of
detection. In order to pay more attention to local de-
tailed features, we introduce the self-position atten-
tion module, which lets local features establish links
in contextual relationships, thereby enhancing the net-
work's expression of local features.
As shown in the upper part of Figure 3, given the fea-
ture , it is first fed into a 1x1 convolu-
tional layer to generate , where (
). We then reshape and into
and , where in the
SPA module. Next, is transposed to
, and the matrix multiplication is performed
between and to obtain their relationship, de-
noted by . After softmax layer, we will ob-
tain position attention maps . From the
spatial position， of and of , we respec-
tively get two spatial points {
}
,where ，. We fur-
ther denote the pointwise calculation of ) as
, and have the following equation:
(2)
where represents feature at position i
impact on position j. The more correlated the features
at the two positions are, the larger is. Then, after
multiplying with the matrix
, we sum
them to obtain the correlation of all features in column
i, :
. (3)
Finally, we aggregate the features of all points and re-
shape to get the final output i.e.:
(3)
Finally, we aggregate the features of all points and re-
shape to get the final output Ep ∈ ℝC×H×W, i.e.:
Figure 3
The details of DSA. Features are fed to the self-position attention (SPA, upper) and self-channel attention (SCA, below)
modules, which output and , respectively. They then fuse into features with positional attention and
channel attention. DSA aims to make the network better understand "what to observe”.
where selectively aggregates the context based
on the position attention map. Similar semantic fea-
tures are correlated with each other, thereby improv-
ing intra-class compactness and semantic consistency.
Self-Channel Attention (SCA). Each channel of
high-level semantic features can be considered a par-
ticular class of response, and the different semantic
responses are interrelated [15]. As a consequence, we
add a self-channel attention module to build depend-
encies between channels.
As shown in the lower half of Figure 3, unlike the
SPA module, before computing the relationship be-
tween the two channels, we directly reshape the input
features to obtain and in-
stead of feeding them into the convolutional layer, as
this maintains the relationship between different
channel mappings. is multiplied by its transposed
feature map to produce channel attention feature map
. The subsequent calculation is similar to
equations (2), (3), and (4), the final output
can be obtained by matrix multiplication
of X and , i.e.:
(5)
Finally, in order to fully fuse the local detail features
between position and channel, we sum the features
from these two attention modules to obtain fusion fea-
ture . The DSA module does not add too many
parameters to the network but effectively enhances
the local representation of the features, allowing the
network to better understand "what to observe."
3.3. Cross-Correlation Attention (CCA)
The GAN model concentrates on learning local fea-
tures and uses an upsampling process to produce vis-
ually enhanced images. However, this local-to-global
structure deviates fundamentally from the global gen-
eration pattern of natural images, resulting in GAN-
generated images frequently lacking global features.
Such images may appear indistinguishable to the hu-
man eye, but in reality, they are merely a composite
of uncorrelated local features.
After applying the DSA module, a pair of features that
aggregate local information can be obtained. How-
ever, since there is no correlation between these two
features, neglecting global information could poten-
tially impact the detection results. Consequently, we
integrated the CCA into the pipeline after the DSA
module to establish inter-image links, facilitating de-
tection and classification.
As shown in Figure 4, the features and
are fed into a 1x1 convolution network to adjust the
channel size. The output is then reshaped such that
and ,where ,
,
(Note, in this paper,
. Here we write separately to better illustrate
the processes involved). Denote the relationship be-
tween and as , we can compute
their cross-attention graph maps . Sim-
ilar as SPA module, of and j of , we re-
spectively get two spatial points {
,
}
,where ， and de-
note the pointwise calculation of ) as
. We choose the cosine similarity
(4)
where Ep selectively aggregates the context based on
the position attention map. Similar semantic features
are correlated with each other, thereby improving in-
tra-class compactness and semantic consistency.
Self-Channel Attention (SCA). Each channel of
high-level semantic features can be considered a par-
ticular class of response, and the dierent semantic
responses are interrelated [15]. As a consequence, we
add a self-channel attention module to build depen-
dencies between channels.
As shown in the lower half of Figure 3, unlike the
SPA module, before computing the relationship
between the two channels, we directly reshape the
input features to obtain f3' ∈ ℝC×N and f2' ∈ ℝC×N in-
stead of feeding them into the convolutional layer,
as this maintains the relationship between dierent
channel mappings. f3' is multiplied by its transposed
feature map to produce channel attention feature
map X ∈ ℝC×C. The subsequent calculation is simi-
lar to equations (2), (3), and (4), the final output EC
∈ ℝC×H×W can be obtained by matrix multiplication of
X and f2', i.e.:
Figure 3
The details of DSA. Features are fed to the self-position attention (SPA, upper) and self-channel attention (SCA, below)
modules, which output and , respectively. They then fuse into features with positional attention and
channel attention. DSA aims to make the network better understand "what to observe”.
where selectively aggregates the context based
on the position attention map. Similar semantic fea-
tures are correlated with each other, thereby improv-
ing intra-class compactness and semantic consistency.
Self-Channel Attention (SCA). Each channel of
high-level semantic features can be considered a par-
ticular class of response, and the different semantic
responses are interrelated [15]. As a consequence, we
add a self-channel attention module to build depend-
encies between channels.
As shown in the lower half of Figure 3, unlike the
SPA module, before computing the relationship be-
tween the two channels, we directly reshape the input
features to obtain and in-
stead of feeding them into the convolutional layer, as
this maintains the relationship between different
channel mappings. is multiplied by its transposed
feature map to produce channel attention feature map
. The subsequent calculation is similar to
equations (2), (3), and (4), the final output
can be obtained by matrix multiplication
of X and , i.e.:
(5)
Finally, in order to fully fuse the local detail features
between position and channel, we sum the features
from these two attention modules to obtain fusion fea-
ture . The DSA module does not add too many
parameters to the network but effectively enhances
the local representation of the features, allowing the
network to better understand "what to observe."
3.3. Cross-Correlation Attention (CCA)
The GAN model concentrates on learning local fea-
tures and uses an upsampling process to produce vis-
ually enhanced images. However, this local-to-global
structure deviates fundamentally from the global gen-
eration pattern of natural images, resulting in GAN-
generated images frequently lacking global features.
Such images may appear indistinguishable to the hu-
man eye, but in reality, they are merely a composite
of uncorrelated local features.
After applying the DSA module, a pair of features that
aggregate local information can be obtained. How-
ever, since there is no correlation between these two
features, neglecting global information could poten-
tially impact the detection results. Consequently, we
integrated the CCA into the pipeline after the DSA
module to establish inter-image links, facilitating de-
tection and classification.
As shown in Figure 4, the features and
are fed into a 1x1 convolution network to adjust the
channel size. The output is then reshaped such that
and ,where ,
,
(Note, in this paper,
. Here we write separately to better illustrate
the processes involved). Denote the relationship be-
tween and as , we can compute
their cross-attention graph maps . Sim-
ilar as SPA module, of and j of , we re-
spectively get two spatial points {
,
}
,where ， and de-
note the pointwise calculation of ) as
. We choose the cosine similarity
(5)
Finally, in order to fully fuse the local detail features
between position and channel, we sum the features
from these two attention modules to obtain fusion
feature ADual. The DSA module does not add too many
parameters to the network but eectively enhances
the local representation of the features, allowing the
network to better understand “what to observe.”
3.3. Cross-Correlation Attention (CCA)
The GAN model concentrates on learning local fea-
tures and uses an upsampling process to produce vi-
sually enhanced images. However, this local-to-glob-
al structure deviates fundamentally from the global
generation pattern of natural images, resulting in
GAN-generated images frequently lacking global fea-
tures. Such images may appear indistinguishable to
the human eye, but in reality, they are merely a com-
posite of uncorrelated local features.
After applying the DSA module, a pair of features that
aggregate local information can be obtained. How-
ever, since there is no correlation between these two
features, neglecting global information could poten-
tially impact the detection results. Consequently, we
integrated the CCA into the pipeline after the DSA
module to establish inter-image links, facilitating de-
tection and classification.
Figure 4
The details of CCA. Aim to adjust the “focus” of the image
given during the network test
As shown in Figure 4, the features fsDual and fqDual are
fed into a 1x1 convolution network to adjust the chan-
nel size. The output is then reshaped such that f1' ∈
ℝC'×N1 and f2' ∈ ℝC×N2 ,where C'< C, N1 = H1× W1, N2 =
H2× W2 (Note, in this paper, N1 = N2. Here we write
separately to better illustrate the processes involved).
Denote the relationship between f1' and f2' as g(f1', f2'),
we can compute their cross-attention graph maps
A∈ℝN1×N2. Similar as SPA module, i of f1' and j of f2', we
respectively get two spatial points {f1i
', f2j
'} ∈ ℝC', where
i ∈ {1, ..., N1}, j ∈ {1, ..., N2} and denote the pointwise
calculation of g(f1', f2') as gij(f1i
', f2j
'). We choose the co-
Information Technology and Control 2024/2/53
414
sine similarity function to calculate the relationship
between features:
Figure 4
The details of CCA. Aim to adjust the "focus" of the
image given during the network test.
function to calculate the relationship between fea-
tures:
where means the cosine similarity between
two features and
.We
perform l2-normalization over and , along
their channel dimension, then equation (6) can be re-
written as:
(7)
The feature map contains the cross-correlation be-
tween each position in A and B. After obtaining the
cross-feature map, it is multiplied by the correspond-
ing matrices of and ,respectively:
. (8)
It can be seen that the output feature
contains
global information of for each pixel, and
is the same. CCA promotes the generation of more
discriminative features for semantically similar re-
gions between support and query images, allowing the
network to adjust its "focus" on the images during
testing. Finally, the channels of the features are ad-
justed to output features
and
.
4. Experiments
In this section, we first introduce the details of the da-
taset and experiments. Then we conduct a series of
contrast experiments and ablation experiments to an-
alyze the effectiveness of the pro-posed model and
modules. Finally, we discuss how to design the RENet
model structure to achieve the best performance of the
network. Additionally, we explore the detection effec-
tiveness across various generated image categories,
including food, animals, landscapes, and more.
4.1 Datasets
In the experiment, the real faces CelebA-HQ [44] and
FFHQ [31] are used as positive samples, and the fake
faces generated by PGGAN [30] and StyleGAN [31]
are used as negative samples, i.e. the number of
categories N = 4. Following [4], we randomly select
10K images from each of the four datasets and then
divide the images into training, validation, and testing
sets in a ratio of 7:1:2. Each image is resized to 256 ×
256. Data augmentation proves to be effective in mit-
igating the overfitting problem in deep learning mod-
els and enhancing their generalization ability [60, 70].
Consequently, during training, we exclusively aug-
ment query images to simulate real-world scenarios.
To assess generalization ability, we randomly select
2,000 images from commonly used out-of-sample da-
tasets as test sets including StyleGAN2 [33], Star-
GAN [11], BeGAN [2], LsGAN [46], WgGANGP
[20], RelGAN [42] The details of the corresponding
generated faces are provided in Table 1 and each im-
age resized to 256 × 256.
Table 1
Details of datasets.
Source Models Resolution
CelebA
StarGAN 256x256
BeGAN 128x128
LsGAN 128x128
WgGANGP 128x128
CelebA-HQ ProGAN 1024x1024
RelGAN 256x256
FFHQ StyleGAN 256x256
StyleGAN2 1024x1024
4.2. Implementation Details
Network architectures. The complete network is
partitioned into three segments: Base Representation,
Feature Augmentation, and Representation Compari-
son. In the Base Representation, we discard the final
average pooling layer and fully connected layer of
ResNet50 [25] and use the remaining layers as the
shared embedding network in the RENet. The struc-
tures and parameters of the remaining two segments
will be thoroughly outlined in Table 2. In Figure 2,
image is input into the shared em-
bedding network to obtain the feature
. In the score network, the features are first
fed into two identical convolutional groups. Each
group contains a 3x3 convolutional layer with 64 fil-
ters, followed by a BN layer, ReLU layer, and max-
pool layer. The final output is transformed to a range
of 0 to 1 using two fully connected layers and a sig-
moid function. The class with the highest score is the
final classification result. In the SPA and CCA mod-
ules, we set 256.
Training and testing details. All experiments are im-
plemented on Pytorch and a GeForce GTX 24GB
3090 GPU, Silver 4214R CPU. The optimizer is
Adam [33]. The initial learning rate of 1.0e-5 and a
decay learning rate of 1.0e-6. Besides a cosine sched-
uler for warm start. Training stops when the learning
(6)
where sim( . , . ) means the cosine similarity between
two features and
Figure 4
The details of CCA. Aim to adjust the "focus" of the
image given during the network test.
function to calculate the relationship between fea-
tures:
two features and
.We
written as:
(7)
The feature map contains the cross-correlation be-
tween each position in A and B. After obtaining the
cross-feature map, it is multiplied by the correspond-
ing matrices of and ,respectively:
. (8)
It can be seen that the output feature
contains
global information of for each pixel, and
is the same. CCA promotes the generation of more
discriminative features for semantically similar re-
gions between support and query images, allowing the
network to adjust its "focus" on the images during
testing. Finally, the channels of the features are ad-
justed to output features
and
.
4. Experiments
In this section, we first introduce the details of the da-
taset and experiments. Then we conduct a series of
contrast experiments and ablation experiments to an-
alyze the effectiveness of the pro-posed model and
modules. Finally, we discuss how to design the RENet
model structure to achieve the best performance of the
network. Additionally, we explore the detection effec-
tiveness across various generated image categories,
including food, animals, landscapes, and more.
4.1 Datasets
In the experiment, the real faces CelebA-HQ [44] and
FFHQ [31] are used as positive samples, and the fake
faces generated by PGGAN [30] and StyleGAN [31]
are used as negative samples, i.e. the number of
categories N = 4. Following [4], we randomly select
10K images from each of the four datasets and then
divide the images into training, validation, and testing
sets in a ratio of 7:1:2. Each image is resized to 256 ×
256. Data augmentation proves to be effective in mit-
igating the overfitting problem in deep learning mod-
els and enhancing their generalization ability [60, 70].
Consequently, during training, we exclusively aug-
ment query images to simulate real-world scenarios.
To assess generalization ability, we randomly select
2,000 images from commonly used out-of-sample da-
tasets as test sets including StyleGAN2 [33], Star-
GAN [11], BeGAN [2], LsGAN [46], WgGANGP
[20], RelGAN [42] The details of the corresponding
generated faces are provided in Table 1 and each im-
age resized to 256 × 256.
Table 1
Details of datasets.
Source Models Resolution
CelebA
StarGAN 256x256
BeGAN 128x128
LsGAN 128x128
WgGANGP 128x128
CelebA-HQ ProGAN 1024x1024
RelGAN 256x256
FFHQ StyleGAN 256x256
StyleGAN2 1024x1024
4.2. Implementation Details
Network architectures. The complete network is
partitioned into three segments: Base Representation,
Feature Augmentation, and Representation Compari-
son. In the Base Representation, we discard the final
average pooling layer and fully connected layer of
ResNet50 [25] and use the remaining layers as the
shared embedding network in the RENet. The struc-
tures and parameters of the remaining two segments
will be thoroughly outlined in Table 2. In Figure 2,
image is input into the shared em-
bedding network to obtain the feature
. In the score network, the features are first
fed into two identical convolutional groups. Each
group contains a 3x3 convolutional layer with 64 fil-
ters, followed by a BN layer, ReLU layer, and max-
pool layer. The final output is transformed to a range
of 0 to 1 using two fully connected layers and a sig-
moid function. The class with the highest score is the
final classification result. In the SPA and CCA mod-
ules, we set 256.
Training and testing details. All experiments are im-
plemented on Pytorch and a GeForce GTX 24GB
3090 GPU, Silver 4214R CPU. The optimizer is
Adam [33]. The initial learning rate of 1.0e-5 and a
decay learning rate of 1.0e-6. Besides a cosine sched-
uler for warm start. Training stops when the learning
.We perform
l2-normalization over f1' and f2', along their channel
dimension, then equation (6) can be rewritten as:
Figure 4
The details of CCA. Aim to adjust the "focus" of the
image given during the network test.
function to calculate the relationship between fea-
tures:
where means the cosine similarity between
two features and
.We
perform l2-normalization over and , along
their channel dimension, then equation (6) can be re-
written as:
(7)
The feature map contains the cross-correlation be-
tween each position in A and B. After obtaining the
cross-feature map, it is multiplied by the correspond-
ing matrices of and ,respectively:
. (8)
It can be seen that the output feature
contains
global information of for each pixel, and
is the same. CCA promotes the generation of more
discriminative features for semantically similar re-
gions between support and query images, allowing the
network to adjust its "focus" on the images during
testing. Finally, the channels of the features are ad-
justed to output features
and
.
4. Experiments
In this section, we first introduce the details of the da-
taset and experiments. Then we conduct a series of
contrast experiments and ablation experiments to an-
alyze the effectiveness of the pro-posed model and
modules. Finally, we discuss how to design the RENet
model structure to achieve the best performance of the
network. Additionally, we explore the detection effec-
tiveness across various generated image categories,
including food, animals, landscapes, and more.
4.1 Datasets
In the experiment, the real faces CelebA-HQ [44] and
FFHQ [31] are used as positive samples, and the fake
faces generated by PGGAN [30] and StyleGAN [31]
are used as negative samples, i.e. the number of
categories N = 4. Following [4], we randomly select
10K images from each of the four datasets and then
divide the images into training, validation, and testing
sets in a ratio of 7:1:2. Each image is resized to 256 ×
256. Data augmentation proves to be effective in mit-
igating the overfitting problem in deep learning mod-
els and enhancing their generalization ability [60, 70].
Consequently, during training, we exclusively aug-
ment query images to simulate real-world scenarios.
To assess generalization ability, we randomly select
2,000 images from commonly used out-of-sample da-
tasets as test sets including StyleGAN2 [33], Star-
GAN [11], BeGAN [2], LsGAN [46], WgGANGP
[20], RelGAN [42] The details of the corresponding
generated faces are provided in Table 1 and each im-
age resized to 256 × 256.
Table 1
Details of datasets.
Source Models Resolution
CelebA
StarGAN 256x256
BeGAN 128x128
LsGAN 128x128
WgGANGP 128x128
CelebA-HQ ProGAN 1024x1024
RelGAN 256x256
FFHQ StyleGAN 256x256
StyleGAN2 1024x1024
4.2. Implementation Details
Network architectures. The complete network is
partitioned into three segments: Base Representation,
Feature Augmentation, and Representation Compari-
son. In the Base Representation, we discard the final
average pooling layer and fully connected layer of
ResNet50 [25] and use the remaining layers as the
shared embedding network in the RENet. The struc-
tures and parameters of the remaining two segments
will be thoroughly outlined in Table 2. In Figure 2,
image is input into the shared em-
bedding network to obtain the feature
. In the score network, the features are first
fed into two identical convolutional groups. Each
group contains a 3x3 convolutional layer with 64 fil-
ters, followed by a BN layer, ReLU layer, and max-
pool layer. The final output is transformed to a range
of 0 to 1 using two fully connected layers and a sig-
moid function. The class with the highest score is the
final classification result. In the SPA and CCA mod-
ules, we set 256.
Training and testing details. All experiments are im-
plemented on Pytorch and a GeForce GTX 24GB
3090 GPU, Silver 4214R CPU. The optimizer is
Adam [33]. The initial learning rate of 1.0e-5 and a
decay learning rate of 1.0e-6. Besides a cosine sched-
uler for warm start. Training stops when the learning
(7)
The feature map contains the cross-correlation be-
tween each position in A and B. After obtaining the
cross-feature map, it is multiplied by the correspond-
ing matrices of f1' and f2',respectively:
Figure 4
The details of CCA. Aim to adjust the "focus" of the
image given during the network test.
function to calculate the relationship between fea-
tures:
where means the cosine similarity between
two features and
.We
perform l2-normalization over and , along
their channel dimension, then equation (6) can be re-
written as:
(7)
The feature map contains the cross-correlation be-
tween each position in A and B. After obtaining the
cross-feature map, it is multiplied by the correspond-
ing matrices of and ,respectively:
. (8)
It can be seen that the output feature
contains
global information of for each pixel, and
is the same. CCA promotes the generation of more
discriminative features for semantically similar re-
gions between support and query images, allowing the
network to adjust its "focus" on the images during
testing. Finally, the channels of the features are ad-
justed to output features
and
.
4. Experiments
In this section, we first introduce the details of the da-
taset and experiments. Then we conduct a series of
contrast experiments and ablation experiments to an-
alyze the effectiveness of the pro-posed model and
modules. Finally, we discuss how to design the RENet
model structure to achieve the best performance of the
network. Additionally, we explore the detection effec-
tiveness across various generated image categories,
including food, animals, landscapes, and more.
4.1 Datasets
In the experiment, the real faces CelebA-HQ [44] and
FFHQ [31] are used as positive samples, and the fake
faces generated by PGGAN [30] and StyleGAN [31]
are used as negative samples, i.e. the number of
categories N = 4. Following [4], we randomly select
10K images from each of the four datasets and then
divide the images into training, validation, and testing
sets in a ratio of 7:1:2. Each image is resized to 256 ×
256. Data augmentation proves to be effective in mit-
igating the overfitting problem in deep learning mod-
els and enhancing their generalization ability [60, 70].
Consequently, during training, we exclusively aug-
ment query images to simulate real-world scenarios.
To assess generalization ability, we randomly select
2,000 images from commonly used out-of-sample da-
tasets as test sets including StyleGAN2 [33], Star-
GAN [11], BeGAN [2], LsGAN [46], WgGANGP
[20], RelGAN [42] The details of the corresponding
generated faces are provided in Table 1 and each im-
age resized to 256 × 256.
Table 1
Details of datasets.
Source Models Resolution
CelebA
StarGAN 256x256
BeGAN 128x128
LsGAN 128x128
WgGANGP 128x128
CelebA-HQ ProGAN 1024x1024
RelGAN 256x256
FFHQ StyleGAN 256x256
StyleGAN2 1024x1024
4.2. Implementation Details
Network architectures. The complete network is
partitioned into three segments: Base Representation,
Feature Augmentation, and Representation Compari-
son. In the Base Representation, we discard the final
average pooling layer and fully connected layer of
ResNet50 [25] and use the remaining layers as the
shared embedding network in the RENet. The struc-
tures and parameters of the remaining two segments
will be thoroughly outlined in Table 2. In Figure 2,
image is input into the shared em-
bedding network to obtain the feature
. In the score network, the features are first
fed into two identical convolutional groups. Each
group contains a 3x3 convolutional layer with 64 fil-
ters, followed by a BN layer, ReLU layer, and max-
pool layer. The final output is transformed to a range
of 0 to 1 using two fully connected layers and a sig-
moid function. The class with the highest score is the
final classification result. In the SPA and CCA mod-
ules, we set 256.
Training and testing details. All experiments are im-
plemented on Pytorch and a GeForce GTX 24GB
3090 GPU, Silver 4214R CPU. The optimizer is
Adam [33]. The initial learning rate of 1.0e-5 and a
decay learning rate of 1.0e-6. Besides a cosine sched-
uler for warm start. Training stops when the learning
(8)
It can be seen that the output feature A1
Cross contains
global information of f1 for each pixel, and A2
Cross is
the same. CCA promotes the generation of more dis-
criminative features for semantically similar regions
between support and query images, allowing the net-
work to adjust its “focus” on the images during test-
ing. Finally, the channels of the features are adjusted
to output features A1
Cross ∈ ℝC×H1×W1 and A2
Cross ∈ ℝC×H2×W2.
4. Experiments
In this section, we first introduce the details of the
dataset and experiments. Then we conduct a series
of contrast experiments and ablation experiments to
analyze the eectiveness of the pro-posed model and
modules. Finally, we discuss how to design the RENet
model structure to achieve the best performance of
the network. Additionally, we explore the detection
eectiveness across various generated image catego-
ries, including food, animals, landscapes, and more.
4.1. Datasets
In the experiment, the real faces CelebA-HQ [44] and
FFHQ [31] are used as positive samples, and the fake
faces generated by PGGAN [30] and StyleGAN [31]
are used as negative samples, i.e. the number of cat-
egories N = 4. Following [4], we randomly select 10K
images from each of the four datasets and then divide
the images into training, validation, and testing sets
in a ratio of 7:1:2. Each image is resized to 256 × 256.
Data augmentation proves to be eective in mitigat-
ing the overfitting problem in deep learning models
and enhancing their generalization ability [60, 70].
Consequently, during training, we exclusively aug-
ment query images to simulate real-world scenarios.
To assess generalization ability, we randomly select
2,000 images from commonly used out-of-sample
datasets as test sets including StyleGAN2 [33], Star-
GAN [11], BeGAN [2], LsGAN [46], WgGANGP [20],
RelGAN [42] The details of the corresponding gener-
ated faces are provided in Table 1 and each image re-
sized to 256 × 256.
4.2. Implementation Details
Network architectures. The complete network is
partitioned into three segments: Base Representa-
tion, Feature Augmentation, and Representation
Comparison. In the Base Representation, we discard
the final average pooling layer and fully connected
layer of ResNet50 [25] and use the remaining lay-
ers as the shared embedding network in the RENet.
The structures and parameters of the remaining two
segments will be thoroughly outlined in Table 2. In
Figure 2, image f ∈ ℝ3×256×256 is input into the shared
embedding network to obtain the feature f ∈ ℝ2048×8×8.
In the score network, the features are first fed into
two identical convolutional groups. Each group con-
Table 1
Details of datasets
Source Models Resolution
CelebA
StarGAN 256x256
BeGAN 128x128
LsGAN 128x128
WgGANGP 128x128
CelebA-HQ ProGAN 1024x1024
RelGAN 256x256
FFHQ StyleGAN 256x256
StyleGAN2 1024x1024
415
Information Technology and Control 2024/2/53
Table 2
Network parameters of feature augmentation module and representation comparison
Feature
Augmentation
Name Parameters
AdaptiveAvgPool2d_1 output_size=8
Conv2d_1
out_channels=256, kernel_size=1,
stride=1
Conv2d_2
out_channels=256, kernel_size=1,
stride=1
Matrix_Mulitiply_1
(Conv2d_1,Conv2d_2)
Softmax_1 softmax(Matrix_Mulitiply_1)
Conv2d_3 out_channels =2048, kernel_size=1,
stride=1
Matrix_Mulitiply_2
(Softmax_1, Conv2d_3)
Matrix_Mulitiply_3
(AdaptiveAvgPool2d_1,
AdaptiveAvgPool2d_1)
Softmax_2 softmax(Matrix_Mulitiply_3)
Matrix_Mulitiply_4
(Softmax_2, Matrix_Mulitiply_3)
Add_S, Add_Q (Matrix_Mulitiply_2,
Matrix_Mulitiply_4)
Conv2d_4_1,
Conv2d_4_2 out_channels =256,
kernel_size=1, stride=1
Conv2d_5_1,
Conv2d_5_2 out_channels =256,
kernel_size=1, stride=1
Matrix_Mulitiply_5_1,
Matrix_Mulitiply_5_2 (Conv2d_4_2, Conv2d_5_1)
(Conv2d_4_1, Conv2d_5_2)
Softmax_3_1,
Softmax_3_2
Softmax_3_1(Matrix_Mulitiply_5_1),
Softmax_3_2(Matrix_Mulitiply_5_2)
Matrix_Mulitiply_6_1,
Matrix_Mulitiply_6_2 (Softmax_3_1,Add_S),
(Softmax_3_2,Add_Q)
Representation
Comparison
Conv2d_6 out_channels =64, kernel_size=3,
stride=1, padding=1
MaxPool2d_1
kernel_size=2, stride=2
Conv2d_7
out_channels =64, kernel_size=3,
stride=1, padding=1
MaxPool2d_2
kernel_size=2, stride=2
AdaptiveAvgPool2d_2 output_size=1
Linear_1 out_features=8
Linear_2 out_features=1
tains a 3x3 convolutional layer
with 64 filters, followed by a BN
layer, ReLU layer, and maxpool
layer. The final output is trans-
formed to a range of 0 to 1 using
two fully connected layers and
a sigmoid function. The class
with the highest score is the fi-
nal classification result. In the
SPA and CCA modules, we set C'
= C/8 = 256.
Training and testing details.
All experiments are implement-
ed on Pytorch and a GeForce
GTX 24GB 3090 GPU, Silver
4214R CPU. The optimizer is
Adam [33]. The initial learning
rate of 1.0e-5 and a decay learn-
ing rate of 1.0e-6. Besides a co-
sine scheduler for warm start.
Training stops when the learn-
ing rate is less than 1.0e-7, with
a patience of 7. We use the mean
square error (MSE) loss when
training, where positive sample
labels are 0 and negative sample
labels are 1. For each episode,
we compute with 8 images from
each category, i.e., K = 8. There-
fore, the experiment is a 4-way
8-shot, and a total of 32 images
(4 x 8) are used as a batch for
training. In addition, to test the
robustness, we perform dier-
ent post-processing operations
on the testing set. Details and
experimental results are given
in Section 4.3.
Performance metric. We use
accuracy (ACC) to evaluate the
experimental results, which is
calculated using the formula
rate is less than 1.0e-7, with a patience of 7. We use
the mean square error (MSE) loss when training,
where positive sample labels are 0 and negative sam-
ple labels are 1. For each episode, we compute with 8
images from each category, i.e., K = 8. Therefore, the
experiment is a 4-way 8-shot, and a total of 32 images
(4 x 8) are used as a batch for training. In addition, to
test the robustness, we perform different post-pro-
cessing operations on the testing set. Details and ex-
perimental results are given in Section 4.3.
Performance metric. We use accuracy (ACC) to
using the formula
. Here, TP
Table 2
Network parameters of feature augmentation module
and representation comparison.
4.3. Comparisons with State-of-the-Art Works
In this section, several advanced methods have been
selected for comparison. They are summarized below.
· Xception [12] performed well in GAN face de-
tection experiments [55]. The optimizer is Adam
[33]. The initial learning rate is set to 1.0e-5, and
the learning rate decay is 1.0e-6. Training stops
when the learning rate is less than 1.0e-7, with a
patience of 7 and batch size is 16. (Xception)
· Yao et al. [65] used transfer learning and feature
fusion module to generated image. The opti-
mizer is SGD optimizer. The model was trained
for 80 epochs with a batch size of 32. (CGNet)
· Ju et al. [29] combined global spatial infor-
mation from the whole image and local informa-
tive features from multiple patches selected by
an adaptation module. The optimizer is Adam
[33]. The initial learning rate is set to 1.0e-5, and
the learning rate decay is 1.0e-6. Training stops
when the learning rate is less than 1.0e-7, with a
patience of 7 and batch size is 32. (FGLNet)
· Li et al. [38] detected authenticity by Estimating
Artifact Similarity. The optimizer is Adam [33].
The initial learning rate is set to 10e-5, then the
model is trained for 200 epochs and optimized
with Adam. (GaseNet)
· Chen et al. [7] proposed a method based on dual-
color space to detect differences. The optimizer
is Adam [33]. The initial learning rate is set to
1.0e-5, and the learning rate decay is 1.0e-6.
Training stops when the learning rate is less than
1.0e-7, with a patience of 7 and batch size is 16.
(Dual-Color)
We mix positive samples from different domains into
one class and negative samples as well, i.e., CelebA-
HQ and FFHQare mixed as one class of positive sam-
ples, and PGGAN and StyleGAN are mixed as one
class of negative samples, and then binary detection is
performed.
Results on original images. As shown in Table 3,
in order to verify the in-sample datasets accuracy and
generalization of the detection methods, we conduct
comparative experiments on different GAN-gener-
ated face datasets (without any post-processing of the
images). It can be seen that the proposed RENet out-
performs the previous methods in both inner and out-
of-sample datasets. RENet benefits from the cross-
image relations of relational network, enabling better
adaptive recognition of image relevance based on
given sample examples. Additionally, it accurately
determines the authenticity of images by contrasting
the attentional differences between images. In Section
4.4, we will provide more detailed results of ablation
and comparative experiments. Although GaseNet
[38] is based on a relational network for detecting ar-
tifacts, the detection performance of RENet is better
than GaseNet. The ROC curves of StyleGAN2 in Fig-
ure 5 also indicate that RENet can maintain excellent
discriminative capability at different decision thresh-
olds.
Furthermore, to better assess the performance of the
models, we compared the inference times of each al-
gorithm. As shown in Table 4, while FGLNet per-
forms well in generalization, its inference time is
Feature
Augmentation
Name Parameters
AdaptiveAvgPool2d_1
output_size=8
Conv2d_1 out_channels =256, kernel_size=1,
stride=1
Conv2d_2 out_channels =256, kernel_size=1,
stride=1
Matrix_Mulitiply_1 (Conv2d_1, Conv2d_2)
Softmax_1 softmax (Matrix_Mulitiply_1)
Conv2d_3 out_channels =2048, kernel_size=1,
stride=1
Matrix_Mulitiply_2 (Softmax_1, Conv2d_3)
Matrix_Mulitiply_3 (AdaptiveAvgPool2d_1,
AdaptiveAvgPool2d_1)
Softmax_2 softmax (Matrix_Mulitiply_3)
Matrix_Mulitiply_4 (Softmax_2, Matrix_Mulitiply_3)
Add_S, Add_Q (Matrix_Mulitiply_2,
Matrix_Mulitiply_4)
Conv2d_4_1,
Conv2d_4_2
out_channels =256, kernel_size=1,
stride=1
Conv2d_5_1,
Conv2d_5_2
out_channels =256, kernel_size=1,
stride=1
Matrix_Mulitiply_5_1,
Matrix_Mulitiply_5_2
(Conv2d_4_2, Conv2d_5_1)
(Conv2d_4_1, Conv2d_5_2)
Softmax_3_1,
Softmax_3_2
Softmax_3_1( Matrix_Mulitiply_5_1) ,
Softmax_3_2( Matrix_Mulitiply_5_2)
Matrix_Mulitiply_6_1,
Matrix_Mulitiply_6_2
(Softmax_3_1, Add_S,)
(Softmax_3_2, Add_Q)
Representation
Comparison
Conv2d_6 out_channels =64, kernel_size=3,
stride=1,
padding=1
MaxPool2d_1 k ernel_size=2, stride=2
Conv2d__7 out_channels =64, kernel_size=3,
stride=1,
padding=1
MaxPool2d_2 k ernel_size=2, stride=2
AdaptiveAvgPool2d_2
output_size=1
Linear_1 out_features=8
Linear_2 out_features=1
. Here,
TP and TN are the numbers of
correctly classified positive and
negative samples, respectively.
P and N are the total number of
positive and negative samples.
Information Technology and Control 2024/2/53
416
4.3. Comparisons with State-of-the-Art Works
In this section, several advanced methods have been
selected for comparison. They are summarized below.
_ Xception [12] performed well in GAN face detec-
tion experiments [55]. The optimizer is Adam [33].
The initial learning rate is set to 1.0e-5, and the
learning rate decay is 1.0e-6. Training stops when
the learning rate is less than 1.0e-7, with a patience
of 7 and batch size is 16. (Xception)
_ Yao et al. [65] used transfer learning and feature fu-
sion module to generated image. The optimizer is
SGD optimizer. The model was trained for 80 ep-
ochs with a batch size of 32. (CGNet)
_ Ju et al. [29] combined global spatial information
from the whole image and local informative fea-
tures from multiple patches selected by an adap-
tation module. The optimizer is Adam [33]. The
initial learning rate is set to 1.0e-5, and the learning
rate decay is 1.0e-6. Training stops when the learn-
ing rate is less than 1.0e-7, with a patience of 7 and
batch size is 32. (FGLNet)
_ Li et al. [38] detected authenticity by Estimating
Artifact Similarity. The optimizer is Adam [33].
The initial learning rate is set to 10e-5, then the
model is trained for 200 epochs and optimized with
Adam. (GaseNet)
_ Chen et al. [7] proposed a method based on dual-col-
or space to detect dierences. The optimizer is
Adam [33]. The initial learning rate is set to 1.0e-5,
and the learning rate decay is 1.0e-6. Training stops
when the learning rate is less than 1.0e-7, with a pa-
tience of 7 and batch size is 16. (Dual-Color)
We mix positive samples from dierent domains
into one class and negative samples as well, i.e., Cele-
bA-HQ and FFHQare mixed as one class of positive
samples, and PGGAN and StyleGAN are mixed as one
class of negative samples, and then binary detection
is performed.
Results on original images. As shown in Table 3, in
order to verify the in-sample datasets accuracy and
generalization of the detection methods, we conduct
comparative experiments on dierent GAN-gener-
ated face datasets (without any post-processing of
the images). It can be seen that the proposed REN-
et outperforms the previous methods in both inner
and out-of-sample datasets. RENet benefits from the
cross-image relations of relational network, enabling
better adaptive recognition of image relevance based
on given sample examples. Additionally, it accurately
determines the authenticity of images by contrasting
the attentional dierences between images. In Section
4.4, we will provide more detailed results of ablation
and comparative experiments. Although GaseNet [38]
is based on a relational network for detecting artifacts,
the detection performance of RENet is better than
GaseNet. The ROC curves of StyleGAN2 in Figure 5
also indicate that RENet can maintain excellent dis-
criminative capability at dierent decision thresholds.
Furthermore, to better assess the performance of
the models, we compared the inference times of each
algorithm. As shown in Table 4, while FGLNet per-
Table 3
Detection accuracy of the different methods (%), the best results of the experiment in bold
Methods
in-sample datasets out-of-sample datasets
ProGAN StyleGAN StyleGAN2 StarGAN BeGAN LsGAN WgGANGP RelGAN
Xception[57] 98.57 98.70 83.97 53.70 49.37 68.84 80.80 73.27
CGNet[67] 99.85 99.23 92.72 98.46 89.20 91.11 98.47 99.94
FGLNet[68] 99.01 98.85 85.63 99.91 94.31 96.52 99.03 95.68
GaseNet[19] 96.80 96.55 89.38 85.54 77.18 90.51 96.33 93.15
Dual-Color[13] 97.53 97.23 79.83 76.30 54.35 82.80 94.80 92.83
RENet (Proposed) 99.93 99.43 94.83 97.17 93.55 98.20 99.35 99.73
417
Information Technology and Control 2024/2/53
Table 4
Detection Average inference time (seconds) comparison of
six methods for a picture with 256 × 256 resolution
Methods Inference time
Xception [65] 0.026
CGNet [67] 0.029
FGLNet [68] 0.045
GaseNet [19] 0.030
Dual-Color [13] 0.052
RENet (Prposed) 0.032
forms well in generalization, its inference time is in-
creased due to the necessity of adaptively selecting
local images for fusion, resulting in an excessive num-
ber of parameters. CGNet, on the other hand, sacrific-
es some generalization performance to improve over-
all eciency. In contrast, RENet maintains excellent
generalization performance while also preserving
good inference times, making it more favorable for
practical applications on real devices.
Results on robustness against post-processing
operations. In reality, some criminals use GAN to
Figure 5
ROC curve for different methods in StyleGAN2
generate faces maliciously and often post-process
the images to evade detection algorithms. Therefore,
we assessed the robustness of our system to several
common attacks. They are JPEG compression (com-
pression factors of 95, 90, 85, 80, 75, 70), Gaussian
blur (kernel size and standard deviation of [3, 0.5], [3,
1.0], [3, 1.5], [5, 0.5], [5, 1.0], [5, 1.5]), resizing (scaling
factors of 40, 60, 80, 120, 140, 160), Gaussian noise
(standard deviation of 0.01, 0.02, 0.03, 0.04, 0.05), and
Gamma correction (gamma values of 0.4, 0.6, 0.8, 1.2,
1.4, 1.6). Our method’s performance was compared to
other methods on both inner and out-of-sample data-
sets, as shown in Figure 6 and Figure 7, respectively.
The results indicate that, in most cases, while the per-
formance of other methods decreased significantly or
Figure 6
The detection accuracy (%) of different methods on in-sample datasets for various post-processing
Information Technology and Control 2024/2/53
418
Figure 7
The detection accuracy (%) of different methods on out-of-sample datasets for various post-processing operations
became ineective when attacked, our method main-
tained ideal detection performance.
This could be attributed to the fact that, even when
post-processings cause degradation, the extracted
features maintain a similarity to the prototype from
the same category. Reliable detection results can be
achieved through feature relation. Additionally, we
visualized the attention points of RENet when pro-
cessing post-processed images. As shown in Figure 8,
it indicates that when the image is disturbed, RENet
not only focuses on the intended areas but also pays
attention to some background edges to assist in dis-
419
Information Technology and Control 2024/2/53
Figure 8
Activation maps from the proposed model after post-
processing operations. From left to right, the column shows
the images after the post-processed images through JPEG
compression, Gaussian blur, resizing, Gaussian noise and
Gamma trans
tinguishing authenticity, aligning with what has been
suggested in [19]. Regarding the suboptimal perfor-
mance of resizing on some datasets, we speculate that
this is due to pixel loss caused by compressing all im-
ages uniformly to the same resolution before training.
4.4. Ablation Study and Selection of RENet
Structure
In this section, we investigate four questions: (1) The
impact of DSA and CCA modules on the eectiveness
of RENet detection. (2) How does the RENet distin-
guish fake faces in out-of-sample datasets? (3) How to
design the network architecture to optimize the per-
formance of RENet. (4) What impact will batch size
and learning rate have on the model?
For the first question, we have conducted ablation ex-
periments in Table 5. The experimental results show
that the DSA and CCA modules can improve the net-
work’s detection generalization ability, thereby verify-
ing the eectiveness of the proposed modules. The vi-
sualization in Figure 9 demonstrates that the features
extracted by the network become more precise and
rational with the addition of the DSA module and CCA
Figure 9
The first column on the left is the input query set, and the
red boxes represent the clearly fake areas. (a), (b), (c) and
(d) are the feature attention maps without any module,
with the DSA module, with the CCA module and with the
DSA+CCA module, respectively
Input (a) (b) (c) (d)
Figure 10
The influence of DSA and CCA on the training loss
Table 5
Ablation experiments of the DSA and CCA modules (%)
ProGAN StyleGAN StyleGAN2 StarGAN BeGAN LsGAN WgGANGP RelGAN
RN 98.6 96.1 93.1 94.2 90.6 97.3 97.6 97.9
RN + DSA 99.9 99.1 94.6 95.7 92.7 98.8 99.3 99.1
RN + CCA 99.5 98.6 93.2 95.4 91.5 93.7 98.6 98.4
RN + DSA + CCA 99.9 99.4 94.8 97.2 93.6 98.2 99.4 99.7
module separately. Especially when combining the two
modules, the network’s attention accuracy further im-
proves. Besides, during the training process, as shown
in Figure 10, the network converges more rapidly after
integrating DSA and CCA. This indicates their success-
ful role in enabling the network to perceive relevant se-
mantic features at dierent positions, and facilitating a
more accessible learning process for comparison.
Information Technology and Control 2024/2/53
420
For the second question, Figure 11 visualizes the pri-
mary focus of RENet on out-of-sample datasets. It
can be observed that, compared to detecting images
within the dataset, when detecting images from out-
of-sample datasets, the network tends to focus on de-
tecting artifacts at the edges of faces and background
hair. It may be due to dierences in the strategies of
dierent generation models, leading to variations in
the correlation of facial texture features.
RENet-1: Instead of element-wise sum, we use ele-
ment-wise avg before feeding the support set features
into the Feature Augmentation. (2) In RENet-2 and
RENet-3, we modify the two convolutional kernels
in the score network to 32 and 128, respectively. (3)
In RENet-4, we add a convolutional layer to the score
network, while in RENet-5, we reduce a convolutional
layer in the score network. To make a fair comparison,
we keep the other structures the same as the original
RENet except for the corresponding modifications.
As shown in Table 6, We can draw the following con-
clusion. First, if we change element-wise sum to ele-
ment-wise avg, the inner and out-of-sample datasets’
detection accuracy of the network will decrease by
0.34% and 1.55%. This is attributed to the fact that el-
ement-wise sum eectively amplifies the distinctions
and commonalities in the extracted features, facil-
itating the network in better contrasting the feature
associations between the support set and query set.
On the contrary, element-wise avg diminishes these
dierences, impacting the network’s ability to discern
their associations. Second, a convolutional kernel of
64 is more suitable for RENet, and a larger (128) or
smaller (32) kernel only brings negative improve-
ments, especially when the convolutional kernel is
128, the generalization accuracy decreases by 3.85%,
indicating that an appropriately sized convolutional
kernel can enhance the feature augmentation results
of DSA. Third, using two convolutional layers before
score mapping is more stable than using one or three
Figure 11
The visualization results of feature maps for out-of-sample
datasets
For the third question, we propose the following veri-
fication directions: (1) After extracting features using
the shared embedding network, should the multiple
features of the support set be element-wise sum or el-
ement-wise avg, which is more suitable for GAN gen-
erated face detection? (2) How many kernels should
the convolutional filtering have in the score network?
(3) How do the convolutional layers in the score net-
work aect detection ability?
To explore the above directions, we have made some
modifications to RENet as shown in Figure 12. (1)
Figure 12
The complete RENet network is on the left, and various modified networks are within the dashed lines on the right. The red font
indicates the modified parameters. (4) and (5) do not modify the parameters but modify the number of convolutional layers
421
Information Technology and Control 2024/2/53
Figure 13
Above are batch size ablation studies, and below are
learning rate ablation studies on the RENet
Table 6
The influence of RENet structure selection on accuracy (%)
Models Description in-sample datasets
average accuracy
out-of-sample datasets
average accuracy
Proposed
RENet \ 99.68 96.85
RENet-1 change support fusion method
into average 99.35 (-0.34) 95.30 (-1.55)
RENet-2 change Score Network’s
kernels into 32 99.58 (-0.1) 95.29 (-1.56)
Modified
RENet-3 change Score Network’s
kernels into 128 98.99 (-0.69) 93.00 (-3.85)
RENet-4 add one same conv in Score
Network 99.06 (-0.62) 96.20 (-0.65)
RENet-5 subtract one same conv in
Score Network 99.11 (-0.57) 94.10 (-2.75)
convolutional layers. Adding a convolutional layer not
only reduces the detection accuracy, but also increas-
es the network parameters, while reducing a layer
cannot fully integrate the extracted features, among
which the impact on generalization accuracy is the
greatest.
For the fourth question, we conduct two experiments.
The first one involved training the model with batch
sizes of 1, 2, 4, 8 (in this paper), and 16, observing the
accuracy on in-sample and out-of-sample datasets
when the epoch equals 20. The second experiment
set the initial learning rates to 10e-4, 10e-5 (in this
paper), and 10e-6, monitoring the accuracy on the
validation set at the same step. As shown in Figures
13 , appropriate batch sizes and learning rates can en-
hance detection accuracy, while excessively small or
large batch sizes may impact generalization ability.
4.5. Generated Image Detection of Various
Category
With the continuous evolution and application of
datasets [5, 61], as well as the ongoing iterations of
GAN models [3, 53, 55, 73, 74], the range of generated
images has extended beyond faces. Training solely on
CelebA-HQ and FFHQ may not comprehensively test
the performance of the proposed method. To thor-
oughly analyze RENet’s performance across dierent
categories, we further expand the testing scope, eval-
uating our proposed method. We employ the experi-
mental settings outlined by Wang et al. [60], which
crafted to show the generality of a generated image
detector trained by a special GAN model in identify-
ing other GAN models (not restricted to faces alone).
Detailed information regarding the generated imag-
Information Technology and Control 2024/2/53
422
Figure 14
Some samples in the experimental datasets. The first row represents real images, and the second row corresponds to
images generated by GAN
Family Models Source Categories
Uncontiditonal GANs
ProGAN LSUN bottle, airplane, chair, horse, etc.
StyleGAN LSUN car, cat, bedroom
StyleGAN2 LSUN car, cat, church, horse
BigGAN ImageNet fish, snake, person, road, etc.
Contiditonal GANs
CycleGAN Style/object transfer apple, orange, zebra, winter, etc.
StarGAN CelebA person
GauGAN COCO bear, truck, spoon, sandwich, etc.
Deepfake FaceForensics++ Videos of faces face
es is presented in Figure 14 and Table 7. We choose
Lsun [66] as the source for real images and select one
category (horse) among the 20 classes generated by
ProGAN as the fake images. Each category comprises
18,000 fake images and an equal number of real imag-
es. The training settings are the same as those men-
tioned in Section 4. In addition to Accuracy (ACC.),
we also use the Average Precision (A.P.) as an evalua-
tion metric. The A.P. is calculated using an alternative
measurement method in method [60], which approx-
imates the area under the precision-recall curve with
the use of a few thresholds.
We first assessed the detection performance of each
method across various categories within the in-sample
datasets. As illustrated in Figure 15, even with training
limited to a single category, our proposed RENet sur-
passes other methods in detecting unseen categories.
Notably, in categories like Bus and Motorbike, where
alternative methods show lower accuracy, RENet
maintains an accuracy exceeding 90%. This suggests
that proposed network is capable of eectively finding
feature connections and semantic information within
the same GAN model. Additionally, Figure 16 visual-
izes the attention of the proposed model on dierent
types of images generated by the same GAN. We ob-
served that, in this scenario, the model primarily fo-
cuses on discerning the authenticity by analyzing the
content within the objects themselves.
Table 8 shows the comparison results between our
proposed RENet and other methods. It can be ob-
served that RENet demonstrates better generaliza-
tion capabilities across most GAN models compared
Table 7
Details of datasets. Generated images are not limited to faces alone
423
Information Technology and Control 2024/2/53
Figure 15
Comparison of performance in unknown categories on images generated by ProGAN(%)
Figure 16
Activation maps in various categories after training on horse images generated by ProGAN
Table 8
Classification accuracy (%) with cross-model & unknow category
Methods
Test Models
ProGAN StyleGAN StyleGAN2 BigGAN CycleGAN StarGAN GauGAN Deepfake Mean
ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P. ACC. A.P.
Xception 84.3 89.5 74.3 80.1 77.3 87.6 58.1 57. 8 60.4 67.1 98.7 99.9 58.0 59.0 54.9 59.0 70.8 75.0
CGNet 93.0 98.9 76.4 96.8 85.2 94.7 71.5 70.1 74.2 80.9 99.8 99.8 52.7 63.1 59.0 70.1 77.2 84.3
FGLNet 87.2 96.1 76.7 85.4 78.4 90.3 64.1 62.9 63.1 70.8 98.1 99.9 48.1 46.3 67.9 72.9 72.9 77.2
GaseNet 86.5 91.2 72.6 77.4 82.5 80.9 68.0 69.1 71.4 71.1 91.5 94.0 61.4 68.8 62.8 66.2 74.6 77.3
Dual-Color 83.3 95.1 70.1 83.8 77.5 85.5 58.2 60.6 63.5 72.3 84.3 95.6 56.3 55.1 57. 8 62.1 68.9 76.3
RENet 94.9 96.2 78.4 81.8 81.6 84.4 71.6 72.2 76.0 80.4 97.8 98.7 63.2 60.5 69.3 69.5 78.1 80.4
Information Technology and Control 2024/2/53
424
to other algorithms. This is attributed to the enhanced
relational network’s ability to perceive semantic fea-
tures and conduct comparisons. However, the gener-
alization performance on GauGAN and Deepfake is
not satisfactory. This happens due to GauGAN and
ProGAN have similar semantic features, leading to
overfitting problems in previous methods. In addi-
tion, the poor performance of the Deepfake model
is due to the fact that it is not a GAN model and uses
MSE loss and SSIM loss for training, resulting in a de-
tection accuracy of only 69.3%.
5. Conclusion
In this work, we propose a relational embedding net-
work called RENet for detecting GAN-generated face.
It combines dual self-attention and cross-attention,
enhancing both the relevant local features within an
image and the global feature relationships between im-
ages. In addition, we observe that RENet can better gen-
eralize to unknown datasets by learning the structural
correlations among features, and bring performance
improvements to the network for detecting GAN-gen-
erated images. In further work, we will explore the ef-
fect of a relational network combined with an attention
framework on dierent image forensics tasks.
Declarations
The authors declare that they have no known compet-
ing interests or personal relationships that could have
appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to thank the anonymous re-
viewers for their valued comments and suggestions.
This work was supported by the projects of nation-
al social science foundation of China under Grant
21BXW077.
Data and Code Availability
The datasets that support the conclusions of this
paper are cited in the article. These datasets were
derived from the following public domain resourc-
es: https://github.com/NVlabs/q-dataset; http://
mmlab.ie.cuhk.edu.hk/projects/CelebA.html; https://
github.com/tkarras/progressive_growing_of_gans.
References
1. Barni, M., Kallas, K., Nowroozi, E., Tondi, B. CNN De-
tection of GAN-Generated Face Images based on
Cross-Band Co-occurrences Analysis. In: 2020 IEEE
International Workshop on Information Fo-rensics
and Security (WIFS). Presented at the 2020 IEEE In-
ternational Workshop on Infor-mation Forensics and
Security (WIFS), 2020, 1-6. https://doi.org/10.1109/
WIFS49906.2020.9360905
2. Berthelot, D., Schumm, T., Metz, L. BE-GAN: Boundary
Equilibrium Generative Adversarial Networks, 2017.
https://doi.org/10.48550/arXiv.1703.10717
3. Brock, A., Donahue, J., Simonyan, K. Large Scale GAN
Training for High Fidelity Natural Image Synthesis,
2018. arXiv preprint arXiv:1809.11096.
4. Chai, L., Bau, D., Lim, S.-N., Isola, P. What Makes Fake
Images Detectable? Understanding Proper-ties that
Generalize, in: Vedaldi, A., Bischof, H., Brox, T., Frahm,
J.-M. (Eds.), Computer Vision - ECCV 2020, Lecture
Notes in Computer Science. Springer International Pu-
blishing, Cham, 103-120. https://doi.org/10.1007/978-
3-030-58574-7_7
5. Chen, B., Ju, X., Xiao, B., Ding, W., Zheng, Y., de Al-
buquerque, V.H.C. Locally GAN-generated Face De-
tection Based on an Improved Xception. Information
Sciences, 2021, 572, 16-28. https://doi.org/10.1016/j.
ins.2021.05.006
6. Chen, B., Li, T., Wang, J., Zhao, G. GAN-Generated Face
Detection with Strong Gener-alization Ability Based
on Quaternions (in Chinese). Journal of Computer-Ai-
ded Design & Computer Graphics, 2022, 34, 734-742.
https://doi.org/10.3724/SP.J.1089.2022.19015
7. Chen, B., Liu, X., Zheng, Y., Zhao, G., Shi, Y.-Q. A Ro-
bust GAN-Generated Face Detection Method Based on
Dual-Color Spaces and an Improved Xception. IEEE
Transactions on Circuits and Sys-tems for Video Tech-
nology, 2022, 32, 3527-3538. https://doi.org/10.1109/
TCSVT.2021.3116679
8. Chen, B., Weijin, T., Wang, Y., Zhao, G. Distin-guishing
Between Natural and GAN-Generated Face Images by
Combining Global and Local Features. Chinese Journal
of Electronics, 2022, 31, 59-67. https://doi.org/10.1049/
cje.2020.00.372
425
Information Technology and Control 2024/2/53
9. Chen, S., Yao, T., Chen, Y., Ding, S., Li, J., Ji, R. Local
Relation Learning for Face Forgery Detec-tion. AAAI
Conference on Artificial Intelligence, 2021. https://doi.
org/10.1609/aaai.v35i2.16193
10. Cheng, H., Yang, S., Zhou, J. T., Guo, L., Wen, B. Frequ-
ency Guidance Matters in Few-Shot Learn-ing. Presen-
ted at the Proceedings of the IEEE/CVF International
Conference on Computer Vision, 2023, 11814-11824.
https://doi.org/10.1109/ICCV51070.2023.01085
11. Choi, Y., Choi, M., Kim, M., Ha, J.-W., Kim, S., Choo, J.
StarGAN: Unified Generative Adversari-al Networks
for Multi-domain Image-to-Image Translation. In:
2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition. Pre-sented at the 2018 IEEE/CVF
Conference on Computer Vision and Pattern Recogniti-
on (CVPR), IEEE, Salt Lake City, UT, 2018, 8789-8797.
https://doi.org/10.1109/CVPR.2018.00916
12. Chollet, F. Xception: Deep Learning with Depth-wise
Separable Convolutions, in: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recogniti-
on. Presented at the Proceedings of the IEEE Conferen-
ce on Computer Vision and Pattern Recognition, 2017,
1251-1258. https://doi.org/10.1109/CVPR.2017.195
13. Durall, R., Keuper, M., Keuper, J. Watch Your Up-Con-
volution: CNN Based Generative Deep Neural Ne-
tworks Are Failing to Reproduce Spec-tral Distribu-
tions. In: 2020 IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR). Presented at
the 2020 IEEE/CVF Con-ference on Computer Visi-
on and Pattern Recogni-tion (CVPR), IEEE, Seattle,
WA, USA, 2020, 7887-7896. https://doi.org/10.1109/
CVPR42600.2020.00791
14. Frank, J., Eisenhofer, T., Schönherr, L., Fischer, A., Ko-
lossa, D., Holz, T. Leveraging Frequency Analysis for
Deep Fake Image Recognition. In: Proceedings of the
37th International Conference on Machine Learning.
Presented at the Interna-tional Conference on Machine
Learning (ICML, PMLR), 2020, 3247-3258.
15. Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., Lu, H. Dual
Attention Network for Scene Segmenta-tion. In: 2019
Proceedings of the IEEE/CVF Con-ference on Com-
puter Vision and Pattern Recogni-tion. Presented at
the 2019 Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, IEEE, 2019.
https://doi.org/10.1109/CVPR.2019.00326
16. Fu, T., Xia, M., Yang, G. Detecting GAN-generated Face
Images Via Hybrid Texture and Sensor Noise Based
Features. Multimedia Tools Applications, 2022, 81,
26345-26359. https://doi.org/10.1007/s11042-022-
12661-1
17. Gao, S., Xia, M., Yang, G. Dual-Tree Complex Wavelet
Transform-Based Direction Correlation for Face Forge-
ry Detection. Security and Com-munication Networks,
2021, 1-10. https://doi.org/10.1155/2021/8661083
18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., War-
de-Farley, D., Ozair, S., Courville, A., Bengio, Y., 2014.
Generative Adversarial Nets. In: Advances in Neural
Information Processing Sys-tems. Presented at the
Advances in Neural Infor-mation Processing Systems,
Curran Associates, Inc., 2021.
19. Gragnaniello, D., Cozzolino, D., Marra, F., Poggi, G., Ver-
doliva, L. Are GAN Generated Images Easy to Detect? A
Critical Analysis of the State-of-the-Art. In: 2021 IEEE
International Conference on Multimedia and Expo
(ICME). Presented at the 2021 IEEE International
Confer-ence on Multimedia and Expo (ICME), 2021,
1-6. https://doi.org/10.1109/ICME51207.2021.9428429
20. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Co-
urville, A.C. Improved Training of Wasser-stein GANs.
In: Advances in Neural Information Processing Sys-
tems. Curran Associates, Inc., 2017.
21. Guo, H., Hu, S., Wang, X., Chang, M.-C., Lyu, S. Ro-
bust Attentive Deep Neural Network for De-tecting
GAN-Generated Faces. IEEE Access, 2022, 10, 32574-
32583. https://doi.org/10.1109/ACCESS.2022.3157297
22. Guo, H., Hu, S., Wang, X., Chang, M.-C., Lyu, S. Eyes Tell
All: Irregular Pupil Shapes Reveal GAN-Generated Fa-
ces. In: ICASSP 2022-2022 IEEE International Con-
ference on Acoustics, Speech and Signal Processing
(ICASSP). Present-ed at the ICASSP 2022-2022 IEEE
International Conference on Acoustics, Speech and
Signal Pro-cessing (ICASSP), IEEE, Singapore, Singa-
pore, 2022, 2904-2908. https://doi.org/10.1109/ICAS-
SP43922.2022.9746597
23. Guo, H., Hu, S., Wang, X., Chang, M.-C., Lyu, S. Eyes Tell
All: Irregular Pupil Shapes Reveal GAN-Generated Faces.
In: ICASSP 2022-2022 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP).
Present-ed at the ICASSP 2022-2022 IEEE International
Conference on Acoustics, Speech and Signal Pro-cessing
(ICASSP), IEEE, Singapore, Singapore, 2022, 2904-2908.
https://doi.org/10.1109/ICASSP43922.2022.9746597
24. Guo, Z., Yang, G., Chen, J., Sun, X. Fake Face De-tection
Via Adaptive Manipulation Traces Ex-traction Network.
Computer Vision and Image Understanding, 2021, 204,
103170. https://doi.org/10.1016/j.cviu.2021.103170
25. Hassanin, M., Anwar, S., Radwan, I., Khan, F.S., Mian,
A. Visual Attention Methods in Deep Learning: An
In-Depth Survey, 2022.
Information Technology and Control 2024/2/53
426
26. He, K., Zhang, X., Ren, S., Sun, J. Deep Residual Lear-
ning for Image Recognition. In: 2016 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR).
Presented at the 2016 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), IEEE, Las Ve-
gas, NV, USA, 2016, 770-778. https://doi.org/10.1109/
CVPR.2016.90
27. Hu, S., Li, Y., Lyu, S. Exposing GAN-Generated Fa-
ces Using Inconsistent Corneal Specular High-lights.
In: ICASSP 2021-2021 IEEE International Confe-
rence on Acoustics, Speech and Signal Pro-cessing
(ICASSP). Presented at the ICASSP 2021-2021 IEEE
International Conference on Acoustics, Speech and
Signal Processing (ICASSP), IEEE, Toronto, ON, Ca-
nada, 2021, 2500-2504. https://doi.org/10.1109/ICAS-
SP39728.2021.9414582
28. Huang, H., Wang, Y., Chen, Z., Zhang, Y., Li, Y., Tang, Z.,
Chu, W., Chen, J., Lin, W., Ma, K.-K. Cmua-watermark:
A Cross-Model Universal Ad-versarial Watermark for
Combating Deepfakes. In: Proceedings of the AAAI
Conference on Arti-ficial Intelligence, 2022, 989-997.
https://doi.org/10.1609/aaai.v36i1.19982
29. Jia, G., Zheng, M., Hu, C., Ma, X., Xu, Y., Liu, L., Deng, Y.,
He, R. Inconsistency-Aware Wavelet Dual-Branch Ne-
twork for Face Forgery Detection. IEEE Transactions on
Biometrics, Behavior, and Identity Science, 2021, 3, 308-
319. https://doi.org/10.1109/TBIOM.2021.3086109
30. Ju, Y., Jia, S., Ke, L., Xue, H., Nagano, K., Lyu, S. Fusing
Global and Local Features for Gener-alized AI-Synthe-
sized Image Detection. In: 2022 IEEE International Con-
ference on Image Processing (ICIP). IEEE, 2022, 3465-
3469. https://doi.org/10.1109/ICIP46576.2022.9897820
31. Karras, T., Aila, T., Laine, S., Lehtinen, J. Progres-si-
ve Growing of GANs for Improved Quality, Sta-bi-
lity, and Variation, 2018. https://doi.org/10.48550/
arXiv.1710.10196
32. Karras, T., Laine, S., Aila, T. A Style-Based Gen-erator
Architecture for Generative Adversarial Networks. In:
Proceedings of the IEEE/CVF Con-ference on Compu-
ter Vision and Pattern Recogni-tion. Presented at the
Proceedings of the IEEE/CVF Conference on Compu-
ter Vision and Pattern Recognition, 2019, 4401-4410.
https://doi.org/10.1109/CVPR.2019.00453
33. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen,
J., Aila, T. Analyzing and Improving the Image Quality
of StyleGAN. In: 2020 IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR). Presen-
ted at the 2020 IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR), IEEE, Seatt-
le, WA, USA, 2020, 8107-8116. https://doi.org/10.1109/
CVPR42600.2020.00813
34. Kingma, D.P., Ba, J. Adam: A Method for Sto-chastic Op-
timization, 2017. arXiv Preprint arXiv:1412.6980.
35. Laurinavičius, D., Maskeliūnas, R., Damaševičius, R.
Improvement of Facial Beauty Prediction Using Arti-
ficial Human Faces Generated by Generative Adversa-
rial Network. Cognitive Computing, 2023, 15, 998-1015.
https://doi.org/10.1007/s12559-023-10117-8
36. Le, B.M., Woo, S.S. ADD: Frequency Attention and
Multi-View Based Knowledge Distillation to Detect
Low-Quality Compressed Deepfake Imag-es. AAAI
Conference on Artificial Intelligence, 2021.
37. Li, H., Li, W., Wang, S. Discovering and Incorpo-rating
Latent Target-Domains for Domain Adap-tation. Pa-
ttern Recognition, 2020, 108, 107536. https://doi.or-
g/10.1016/j.patcog.2020.107536
38. Li, J., Xie, H., Jiahong Li, Wang, Z., Zhang, Y. Frequen-
cy-Aware Discriminative Feature Learn-ing Supervised
by Single-Center Loss for Face Forgery Detection. In: 2021
IEEE/CVF Confer-ence on Computer Vision and Pattern
Recognition (CVPR). Presented at the 2021 IEEE/CVF
Con-ference on Computer Vision and Pattern Recogni-ti-
on (CVPR), IEEE, Nashville, TN, USA, 2021, 6454-6463.
https://doi.org/10.1109/CVPR46437.2021.00639
39. Li, W., He, P., Li, H., Wang, H., Zhang, R. Detec-tion of
GAN-Generated Images by Estimating Ar-tifact Simi-
larity. IEEE Signal Processing Letters, 2022, 29, 862-
866. https://doi.org/10.1109/LSP.2021.3130525
40. Li, Wei, Zhong, X., Shao, H., Cai, B., Yang, X. Multi-Mo-
de Data Augmentation and Fault Diag-nosis of Rota-
ting Machinery Using Modified AC-GAN Designed
with New Framework. Ad-vanced Engineering Infor-
matics, 2022, 52, 101552. https://doi.org/10.1016/j.
aei.2022.101552
41. Liang, B., Wang, Z., Huang, B., Zou, Q., Wang, Q., Liang, J.
Depth Map Guided Triplet Network for Deepfake Face
Detection. Neural Networks, 2023, 159, 34-42. https://
doi.org/10.1016/j.neunet.2022.11.031
42. Lin, H., Luo, W., Wei, K., Liu, M. Improved Xcep-tion
with Dual Attention Mechanism and Feature Fusion
for Face Forgery Detection. In: 2022 4th International
Conference on Data Intelligence and Security (ICDIS).
Presented at the 2022 4th Inter-national Conference
on Data Intelligence and Se-curity (ICDIS), 2022, 208-
212. https://doi.org/10.1109/ICDIS55630.2022.00039
43. Lin, Y.-J., Wu, P.-W., Chang, C.-H., Chang, E., Liao, S.-
W. RelGAN: Multi-Domain Im-age-to-Image Trans-
427
Information Technology and Control 2024/2/53
lation via Relative Attributes. In: 2019 IEEE/CVF In-
ternational Conference on Computer Vision (ICCV).
Presented at the 2019 IEEE/CVF International Con-
ference on Computer Vision (ICCV), IEEE, Seoul, Ko-
rea (South), 2019, 5913-5921. https://doi.org/10.1109/
ICCV.2019.00601
44. Liu, H., Li, X., Zhou, W., Chen, Y., He, Y., Hui, X., Wei-
ming, Z., Nenghai, Y. Spatial-Phase Shallow Learning:
Rethinking Face Forgery Detection in Frequency Do-
main. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition. Presented
at the Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2021, 772-
781. https://doi.org/10.1109/CVPR46437.2021.00083
45. Liu, Z., Luo, P., Wang, X., Tang, X. Deep Learning Face
Attributes in the Wild, in: Proceedings of the IEEE In-
ternational Conference on Computer Vi-sion. Presen-
ted at the Proceedings of the IEEE In-ternational Con-
ference on Computer Vision, 2015, 3730-3738. https://
doi.org/10.1109/ICCV.2015.425
46. Liu, Z., Qi, X., Torr, P. H. S. Global Texture En-hancement
for Fake Face Detection in the Wild. In: 2020 IEEE/
CVF Conference on Computer Vi-sion and Pattern Re-
cognition (CVPR). Presented at the 2020 IEEE/CVF
Conference on Computer Vision and Pattern Recogni-
tion (CVPR), IEEE, Seattle, WA, USA, 2020, 8057-8066.
https://doi.org/10.1109/CVPR42600.2020.00808
47. Mao, X., Li, Q., Xie, H., Lau, R. Y., Wang, Z., Paul Smol-
ley, S. Least Squares Generative Adversarial Networks.
In: Proceedings of the IEEE Interna-tional Conferen-
ce on Computer Vision, 2017, 2794-2802. https://doi.
org/10.1109/ICCV.2017.304
48. Matern, F., Riess, C., Stamminger, M. Exploiting Visual
Artifacts to Expose Deepfakes and Face Manipulations.
In: 2019 IEEE Winter Applica-tions of Computer Visi-
on Workshops (WACVW). Presented at the 2019 IEEE
Winter Applications of Computer Vision Workshops
(WACVW), IEEE, Waikoloa Village, HI, USA, 2019, 83-
92. https://doi.org/10.1109/WACVW.2019.00020
49. Miao, C., Tan, Z., Chu, Q., Liu, H., Hu, H., Yu, N. F 2 Trans:
High-Frequency Fine-Grained Trans-former for Face
Forgery Detection. IEEE Transac-tions on Information
Forensics and Security, 2023, 18, 1039-1051. https://doi.
org/10.1109/TIFS.2022.3233774
50. Nataraj, L., Mohammed, T. M., Manjunath, B. S., Chan-
drasekaran, S., Flenner, A., Bappy, J. H., Roy-Chowdhu-
ry, A. K. Detecting GAN Generated Fake Images Using
Co-occurrence Matrices. ei 31, 2019, 532-1-532-7.
https://doi.org/10.2352/ISSN.2470-1173.2019.5.M-
WSF-532
51. Neekhara, P., Hussain, S., Zhang, X., Huang, K., McAu-
ley, J., Koushanfar, F. FaceSigns: Semi-Fragile Neural
Watermarks for Media Au-thentication and Countering
Deepfakes, 2022. arXiv Preprint arXiv:2204.01960.
52. Ouyang, J., Huang, J., Wen, X., Shao, Z. A Semi-Fragile
Watermarking Tamper Localization Method Based on
QDFT and Multi-View Fusion. Multimedia Tools Appli-
cation, 2022. https://doi.org/10.1007/s11042-022-13938-1
53. Pan, X., Ge, C., Lu, R., Song, S., Chen, G., Huang, Z., Hu-
ang, G. On the Integration of Self-Attention and Convo-
lution. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2022, 815-
825. https://doi.org/10.1109/CVPR52688.2022.00089
54. Park, T., Liu, M.-Y., Wang, T.-C., Zhu, J.-Y. Se-man-
tic Image Synthesis with Spatially-Adaptive Norma-
lization. Presented at the Proceedings of the IEEE/
CVF Conference on Computer Vision and Pattern Re-
cognition, 2019, 2337-2346. https://doi.org/10.1109/
CVPR.2019.00244
55. Rhee, H., Jang, Y. I., Kim, S., Cho, N. I. LC-FDNet: Lear-
ned Lossless Image Compression with Fre-quency De-
composition Network. In: 2022 IEEE/CVF Conference
on Computer Vision and Pattern Recognition (CVPR).
Presented at the 2022 IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR), IEEE,
New Or-leans, LA, USA, 2022, 6023-6032. https://doi.
org/10.1109/CVPR52688.2022.00594
56. Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies,
J., Niessner, M. Face Forensics: Learn-ing to Detect Ma-
nipulated Facial Images, in: Pro-ceedings of the IEEE/
CVF International Confer-ence on Computer Vision.
Presented at the Pro-ceedings of the IEEE/CVF Inter-
national Confer-ence on Computer Vision, 2019, 1-11.
https://doi.org/10.1109/ICCV.2019.00009
57. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P. H. S., Hos-
pedales, T. M. Learning to Compare: Rela-tion Network
for Few-Shot Learning. In: 2018 IEEE/CVF Conferen-
ce on Computer Vision and Pattern Recognition. Pre-
sented at the 2018 IEEE/CVF Conference on Compu-
ter Vision and Pattern Recognition (CVPR), IEEE, Salt
Lake City, UT, 2018, 1199-1208. https://doi.org/10.1109/
CVPR.2018.00131
58. Vasudeva B., Deora P., Bhattacharya S., Pradhan P.M.
Compressed Sensing MRI Reconstruction With
Co-VeGAN: Complex-Valued Generative Adversarial
Network. Presented at the Proceed-ings of the IEEE/
CVF Winter Conference on Ap-plications of Com-
puter Vision, 2022, 672-681. https://doi.org/10.1109/
WACV51458.2022.00184
Information Technology and Control 2024/2/53
428
59. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jo-
nes, L., Gomez, A. N., Kaiser, Ł., Polosukhin, I. Attenti-
on is All you Need, in: Advances in Neural Information
Processing Systems. Curran Associ-ates, Inc, 2017.
60. Wang, J., Zeng, K., Ma, B., Luo, X., Yin, Q., Liu, G., Jha, S.
K. GAN-generated Fake Face Detection via Two-stream
CNN with PRNU in the Wild. Multimedia Tools Appli-
cation, 2022, 1-19. https://doi.org/10.1007/s11042-021-
11592-7
61. Wang, S.-Y., Wang, O., Zhang, R., Owens, A., Efros, A. A..
CNN-Generated Images Are Surpris-ingly Easy to Spot
for Now. In: Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition. Presen-
ted at the Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2020, 8695-
8704. https://doi.org/10.1109/CVPR42600.2020.00872
62. Wei, W., Ho, E.S.L., McCay, K. D., Damaševičius, R., Ma-
skeliūnas, R., Esposito, A. Assessing Facial Symmetry
and Attractiveness Using Augmented Reality. Pattern
Analysis and Applications, 2022, 25, 635-651. https://
doi.org/10.1007/s10044-021-00975-z
63. Wu, H., Kuo, H.-C., Zheng, N., Hung, K.-H., Lee, H.-Y.,
Tsao, Y., Wang, H.-M., Meng, H. Partially Fake Audio De-
tection by Self-Attention-Based Fake Span Discovery.
In: ICASSP 2022-2022 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP).
IEEE, 2022, 9236-9240. https://doi.org/10.1109/ICAS-
SP43922.2022.9746162
64. Wu, Z., Li, Y., Guo, L., Jia, K. Parn: Posi-tion-aware re-
lation Networks for Few-shot Learning. In: Proceedings
of the IEEE/CVF Inter-national Conference on Com-
puter Vision, 2019, 6659-6667. https://doi.org/10.1109/
ICCV.2019.00676
65. Yang, X., Li, Y., Lyu, S. Exposing Deep Fakes Using In-
consistent Head Poses. In: ICASSP 2019 - 2019 IEEE
International Conference on Acous-tics, Speech and
Signal Processing (ICASSP). Presented at the ICASSP
2019-2019 IEEE Interna-tional Conference on Aco-
ustics, Speech and Sig-nal Processing (ICASSP), IEEE,
Brighton, United Kingdom, 2019, 8261-8265. https://
doi.org/10.1109/ICASSP.2019.8683164
66. Yao, Y., Zhang, Z., Ni, X., Shen, Z., Chen, L., Xu, D.
CGNet: Detecting Computer-generated Images Based
on Transfer Learning with Attention Mod-ule. Signal
Processing: Image Communication, 2022, 105, 116692.
https://doi.org/10.1016/j.image.2022.116692
67. Yu, F., Se, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.
Lsun: Construction of a large-scale image dataset using
deep learning with humans in the loop, 2015. arXiv pre-
print arXiv:1506.03365.
68. Yuchen, L., Yon, Z., Junchi, Y., Wei, L. General-izing
Face Forgery Detection with High-Frequency Features.
Presented at the Pro-ceedings of the IEEE/CVF Con-
ference on Com-puter Vision and Pattern Recognition,
2021, 16317-16326.
69. Yuyang, Q., Guojun, Y., Lu, S., Zixuan, C., Jing, S. Thin-
king in Frequency: Face Forgery Detection by Mining
Frequency-Aware Clues, in: Andrea Vedaldi, Horst
Bischof, Thomas Brox, Jan-Michael Frahm (Eds.), Eu-
ropean Conference on Computer Vision-ECCV 2020,
Lecture Notes in Computer Science. Springer Inter-
national Pub-lishing, Cham, 2020, 86-103. https://doi.
org/10.1007/978-3-030-58610-2_6
70. Zheng, Q., Tian, X., Yu, Z., Jiang, N., Elhanashi, A.,
Saponara, S., Yu, R. Application of Wave-let-packet
Transform Driven Deep Learning Method in PM2.5
Concentration Prediction: A Case Study of Qingdao,
China. Sustainable Cities and Society, 2023, 92, 104486.
https://doi.org/10.1016/j.scs.2023.104486
71. Zheng, Q., Zhao, P., Li, Y., Wang, H., Yang, Y. Spectrum
Interference-Based Two-Level Data Augmentation Me-
thod in Deep Learning for Au-tomatic Modulation Clas-
sification. Neural Com-puting & Applications, 2021, 33,
7723-7745. https://doi.org/10.1007/s00521-020-05514-1
72. Zheng, Q., Zhao, P., Wang, H., Elhanashi, A., Sa-ponara,
S. Fine-Grained Modulation Classifica-tion Using Mul-
ti-Scale Radio Transformer with Dual-Channel Repre-
sentation. IEEE Communica-tion Letters, 2022, 26, 1298-
1302. https://doi.org/10.1109/LCOMM.2022.3145647
73. Zhong, Y., Li, B., Tang, L., Kuang, S., Wu, S., Ding, S.
Detecting Camouflaged Object in Frequency Domain
In: 2022 IEEE/CVF Conference on Com-puter Vi-
sion and Pattern Recognition (CVPR). Presented at
the 2022 IEEE/CVF Conference on Computer Vision
and Pattern Recognition (CVPR), IEEE, New Orleans,
LA, USA, 2022, 4494-4503. https://doi.org/10.1109/
CVPR52688.2022.00446
74. Zhu, J.-Y., Park, T., Isola, P., Efros, A. A. Unpaired Ima-
ge-to-Image Translation Using Cy-cle-Consistent
Adversarial Networks. In: 2017 IEEE International
Conference on Computer Vi-sion (ICCV). Presented at
the 2017 IEEE Interna-tional Conference on Computer
Vision (ICCV), IEEE, Venice, 2017, 2242-2251. https://
doi.org/10.1109/ICCV.2017.244
75. Zou, H., Ak, K.E., Kassim, A. A. Edge-Gan: Edge Con-
ditioned Multi-View Face Image Generation. In: 2020
IEEE International Conference on Image Processing
(ICIP). Presented at the 2020 IEEE In-ternational Con-
ference on Image Processing (ICIP), 2020, 2401-2405.
https://doi.org/10.1109/ICIP40778.2020.9190723
This article is an Open Access article distributed under the terms and conditions of the Creative
Commons Attribution 4.0 (CC BY 4.0) License (http://creativecommons.org/licenses/by/4.0/).
