Low quality deepfake detection device and method of detecting low quality deepfake using the same

Patent metadata

Field	Value
Publication number	US-12586419-B2
Application number	US-202318241317-A
Country	US
Kind code	B2
Filing date	Sep 1, 2023
Priority date	Sep 1, 2022
Publication date	Mar 24, 2026
Grant date	Mar 24, 2026

Abstract

A deepfake detection device may include a data input unit that receives an input image including a low-quality deepfake video, a branch-based super-resolution training unit that enhances the resolution of the input image through unsupervised super-resolution training and generates a plurality of super-resolution images having different sizes, and a multi-scale training unit that performs multi-scale training, without resolution conversion, on the plurality of super-resolution images having different sizes, respectively. The multi-scale training unit may synthesize multi-scale training results for the plurality of super-resolution images having different sizes, respectively, and determine whether the input image is a deepfake based on the multi-scale training results.

First claim

What is claimed is:

1 . A deepfake detection device, the device comprising:
memory storing instructions; and at least one processor, wherein the instructions, when executed by the at least one processor, cause the device to
receive an input image including a low-quality deepfake video;
enhance the resolution of the input image through unsupervised super-resolution training and
generate a plurality of super-resolution images having different sizes; and
perform multi-scale training, without resolution conversion, on the plurality of super-resolution images having different sizes, respectively;
synthesize multi-scale training results for the plurality of super-resolution images having different sizes, respectively; and
determine whether the input image is a deepfake based on the multi-scale training results.

2 . The device of claim 1 , wherein the instructions, when executed by the at least one processor, further cause the device to:
receive a guide image and generate a super-resolution image for the guide image; compare at least one of the plurality of super-resolution images generated for the input image with the super-resolution image for the guide image to evaluate training performance for the input image; and compare the input image with the at least one of the plurality of super-resolution images generated for the input image and perform transfer training on the at least one of the plurality of super-resolution images.

3 . The device of claim 2 , wherein the instructions, when executed by the at least one processor, further cause the device to:
compare the guide image with the super-resolution image for the guide image based on a L id loss function, and wherein the L id loss function is defined as [Equation 1] below;

ℒ
id

=

1
N

⁢

∑

i
=
1

N



x
^

i
hqsr

-

x
i
hq



2

[

Equation
⁢

1

]

(Here, N is the total number of branches i,

x
^

i
hqsr

is a super-resolution image for a guide image, and

x
i
hq

is a guide image).

4 . The device of claim 2 , wherein the instructions, when executed by the at least one processor, further cause the device to:
evaluate training performance for the input image based on an L adv loss function, and wherein the L adv loss function is defined as [Equation 2] below;

ℒ
adv

=

1
N

⁢

∑

i
=
1

N

log
⁡
(

D
⁡
(

x
i
hq

)

)

+

log
⁡
(

1
-

D
⁡
(

x
^

i
sr

)

)

[

Equation
⁢

2

]

(Here, N is the total number of branches i, D is a discriminator function,

x
^

i
sr

is the at least one of the plurality of super-resolution images generated for the input image, and

x
i
hq

is a guide image).

5 . The device of claim 2 , wherein the instructions, when executed by the at least one processor, further cause the device to:
perform transfer training on the at least one of the plurality of super-resolution images based on an L perc loss function, and wherein the L perc loss function is defined as [Equation 3] below;

Assignees

• Research & Business Found Sungkyunkwan Univ

Inventors

• Woo Simon Sungil
• Lee Sang Yup
• An Jae Ju

Classifications

• G06T3/4053 Primary

  based on super-resolution, i.e. the output image resolution being higher than the sensor resolution · CPC title

• G06T3/4007

  based on interpolation, e.g. bilinear interpolation (image demosaicing G06T3/4015; edge-driven or edge-based scaling G06T3/403) · CPC title

• G06T3/4046

  using neural networks · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title