Deep Forgery Detection Using CutMix Algorithm and Improved Xception Network

Pengzhi Geng; Yunqi Tang; Hongxing Fan; Xintong Zhu

doi:10.3788/LOP202259.1615007

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 16 >Page 1615007 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 16, 1615007 (2022)

Deep Forgery Detection Using CutMix Algorithm and Improved Xception Network

Pengzhi Geng¹, Yunqi Tang^1、*, Hongxing Fan², and Xintong Zhu¹

Author Affiliations

¹School of Criminal Investigation, People’s Public Security University of China, Beijing 100038, China

²Center for Research on Intelligent Perception and Computing, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China

show less

DOI: 10.3788/LOP202259.1615007 Cite this Article Set citation alerts

Pengzhi Geng, Yunqi Tang, Hongxing Fan, Xintong Zhu. Deep Forgery Detection Using CutMix Algorithm and Improved Xception Network[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1615007 Copy Citation Text

show less

Fig. 1. CutMix augmented image example. (a)(b) Original samples; (c) augmented sample

Download full size

Fig. 2. Sampler for unbalanced data sets

Download full size

Fig. 3. Proposed network structure

Download full size

Fig. 4. ROC curves and AUC values of different models on the validation set

Download full size

Fig. 5. Influence of hyper-parameter α and probability p on the detection model. (a) CutMix; (b) Mixup

Download full size

Fig. 6. Results of data enhancement

Download full size

Input size	Operator		Number of channels
299×299×3	Entry flow	Conv1，2×2	32
149×149×32		Conv2，3×3	64
147×147×64		Block1	128
74×74×128		Block2	256
37×37×256		Block3	728
19×19×728	Middle flow	Block4，3×3	728
19×19×728		Block5，3×3	728
19×19×728		Block6，3×3	728
19×19×728		Block7，3×3	728
19×19×728		Block8，3×3	728
19×19×728		Block9，3×3	728
19×19×728		Block10，3×3	728
19×19×728		Block11，3×3	728
19×19×728	Exit flow	Block12	1024
10×10×1024		SeparableConv2d，3×3	1536
10×10×1536		SeparableConv2d，3×3	2048
10×10×2048		Pool，1×1

Table 1. XceptionNet structure

Dataset	Number of fake images	Number of real images
Train dataset	28800	7200
Test dataset	5600	1400
Validation dataset	5600	1400

Table 2. Description of dataset

Description	Logloss	Accuracy	Parameters /10⁶
Block6	0.5539	0.8554	5.95
Block7	0.5386	0.8631	7.56
Block8	0.5258	0.8721	9.18
Block9	0.5185	0.8684	10.79
Block10	0.5398	0.8687	12.41
XceptionNet	0.5497	0.8757	20.81

Table 3. Model optimization experiment

Method	Model	Logloss	Accuracy	Parameters /10⁶
Method in Ref.［18］	EfficientNet_b3	0.5840	0.8803	12.23
Method in Ref.［28］	ResNet50	0.5413	0.8684	25.56
Method in Ref.［10］	SPPNet	0.8092	0.8660	25.64
Method in Ref.［20］	XceptionNet	0.5497	0.8757	20.81
Proposed method	Xcep_Block8	0.5258	0.8721	9.18

Table 4. Comparison between the proposed model and other classical algorithms

Description	Logloss	Accuracy
Xcep_Block8	0.5258	0.8721
Xcep_Block8+Over sampling	0.4329	0.8751
XceptionNet	0.5497	0.8757
XceptionNet+Over sampling	0.4725	0.8779

Table 5. Improvements for the imbalance of sample categories

Description	CutMix		Mixup
Description	Logloss	Accuracy	Logloss	Accuracy
$α$ =0.5，p=0.7	0.3271	0.8760	0.3196	0.8780
$α$ =0.5，p=0.8	0.3270	0.8759	0.3080	0.8724
$α$ =0.5，p=0.9	0.3334	0.8734	0.3117	0.8737
$α$ =0.5，p=1	0.3097	0.8819	0.3170	0.8731
$α$ =1，p=0.7	0.3210	0.8743	0.3151	0.8683
$α$ =1，p=0.8	0.3376	0.8767	0.3162	0.8704
$α$ =1，p=0.9	0.3122	0.8822	0.3147	0.8686
$α$ =1，p=1	0.3185	0.8773	0.3211	0.8667

Table 6. Influence of different parameter settings on hybrid data enhancement results

Description	Logloss	Accuracy
Baseline（Xcep_Block8）	0.4329	0.8751
+Cutout（size is 50）	0.5143	0.8760
+Cutout（size is 80）	0.5228	0.8750
+Cutout（size is 110）	0.4674	0.8744
+Mixup	0.3196	0.8780
+CutMix	0.3122	0.8822

Table 7. Experimental results of data augmentation

Pengzhi Geng, Yunqi Tang, Hongxing Fan, Xintong Zhu. Deep Forgery Detection Using CutMix Algorithm and Improved Xception Network[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1615007

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information