Deepfake Detection Algorithm for High-Frequency Components of Shallow Features

Fig. 6. Effect before and after enhancement of high-frequency feature map corresponding to real and fake samples. (a) (b) Original images $i$ ; (c) (d) high-frequency feature maps $H L_{t} (i)$ before enhancement; (e) (f) enhanced high-frequency feature map $E n [H L_{t} (i)]$

Download full size

Fig. 7. Template for Laplace operator

Download full size

Fig. 8. Feature map variations. (a) Shallow feature maps $F L_{t} (i)$ ; (b) high-frequency feature maps $H L_{t} (i)$ ; (c) enhanced high-frequency feature maps $E n [H L_{t} (i)]$

Download full size

Fig. 9. Spectrogram variations. (a) Shallow feature maps $F L_{t} (i)$ ; (b) high-frequency feature maps $H L_{t} (i)$ ; (c) enhanced high-frequency feature maps $E n [H L_{t} (i)]$

Download full size

Fig. 10. Variation of model Acc index with the number of network layers on 3 datasets

Download full size

Fig. 11. Variation of model AUC index with the number of network layers on 3 datasets

Download full size

Fig. 12. Line graph of change in average Acc index with $λ$

Download full size

Fig. 13. Line graph of change in average AUC index with $λ$

Download full size

Fig. 14. Comparison of average Acc on three datasets after introduction of GC

Download full size

Fig. 15. Comparison of average loss on three datasets after introduction of GC

Download full size

Table 1. Architecture of XceptionNet

View table

Table 1. Architecture of XceptionNet

Layer	Operator		Output，Size	Channels
$L_{0}$	Entry flow	Conv1		32
$L_{0}$		Conv2		64
$L_{1}$		Block1		128
$L_{2}$		Block2	$F L_{2} (i)$ ，（40×40）	256
$L_{3}$		Block3	$F L_{3} (i)$ ，（20×20）	728
$L_{4}$	Middle flow	Block4	$F L_{4} (i)$ ，（20×20）	728
$L_{5}$ - $L_{11}$	Middle flow	Block5-11		728
$L_{12}$	Exit flow	Block12		1024
$F i n a l$		SeparableConv2d		1536
$F i n a l$		SeparableConv2d		2048
$L o g i t s$		Linear

Table 2. Architecture of EfficientNet-B4

View table

Table 2. Architecture of EfficientNet-B4

Layer	Operator	Numbers	Output，Size	Channels
$L_{0}$	Conv 3×3	1		48
$L_{1}$	MBConv1	2		24
$L_{2}$	MBConv6	4	$F L_{2} (i)$ ，（80×80）	32
$L_{3}$	MBConv6	4	$F L_{3} (i)$ ，（40×40）	56
$L_{4}$	MBConv6	6	$F L_{4} (i)$ ，（20×20）	112
$L_{5}$	MBConv6	6		160
$L_{6}$	MBConv6	8		272
$L_{7}$	MBConv6	2		448
$L o g i t s$	Linear	1

Table 3. Datasets used
View table
Table 3. Datasets used
Dataset DeepFakes FaceSwap Celeb-DF
Train 41936 41945 143192
Test 17089 15600 41618

Table 4. Comparison of gains produced by different modules on XceptionNet-based Baseline

View table

Table 4. Comparison of gains produced by different modules on XceptionNet-based Baseline

X_model			DeepFakes		FaceSwap		Celeb-DF
Baseline	CBAM	GD	Acc /%	AUC	Acc /%	AUC	Acc /%	AUC
√			98.67	0.9867	99.33	0.9933	97.49	0.9037
√	√		98.83	0.9883	99.67	0.9967	98.07	0.9342
√		√	98.83	0.9883	99.50	0.9950	98.01	0.9384
√	√	√	99.33	0.9933	99.83	0.9983	98.23	0.9405

Table 5. Comparison of gains produced by different modules on EfficientNet-B4-based Baseline

View table

Table 5. Comparison of gains produced by different modules on EfficientNet-B4-based Baseline

En_model			DeepFakes		FaceSwap		Celeb-DF
Baseline	Cbam	GD	Acc /%	AUC	Acc /%	AUC	Acc /%	AUC
√			98.50	0.9850	99.00	0.9900	97.56	0.9221
√	√		98.83	0.9883	99.50	0.9950	97.86	0.9263
√		√	98.67	0.9867	99.17	0.9917	97.63	0.9220
√	√	√	99.17	0.9917	99.50	0.9950	98.45	0.9485

Table 6. Comparison of Acc index of each algorithm on three datasets

View table

Table 6. Comparison of Acc index of each algorithm on three datasets

Acc /%	DeepFakes	FaceSwap	Celeb-DF
EfficientNet-B4^［11］	98.33	98.83	97.42
XceptionNet^［10］	97.83	98.17	96.97
MesoNet^［14］	95.50	93.33	91.72
Mo et al^［18］	96.67	97.00	95.79
Sabir et al ^［8］	96.50	96.33	95.42
ResNet34^［9］	93.83	94.17	93.50
En_model	99.17	99.50	98.45
X_model	99.33	99.83	98.23

Tools

Get Citation

Copy Citation Text

Shufan Peng, Manchun Cai, Rui Ma, Xiaowen Liu. Deepfake Detection Algorithm for High-Frequency Components of Shallow Features[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1015001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites