Page 11 - IJEEE-2022-Vol18-ISSUE-1

Kadhim & Al-Darraji |7

images with several shots and different angles and attack with different cases. Table 4 illustrates the accuracy of
directions. face recognition for several situations.

Furthermore, they can only be seen by the Viola-Jones ? Case 1: when training the CNN model using original
face detector. The LFW dataset was next processed, and image and testing with original and distorted images.
Figure 13 displays a sample of the post-processing results.
? Case 2: when training the CNN model using original
and distorted images and testing with original images.

? Case 3: when training the CNN model using original
and distorted images and testing with original and
distorted images. However, the proportions are varied
with respect to accuracy.

Fig. 13: Some Face Images of Different Subjects of the LFW As for the final merge experiment, training is done in
Database. three cases:

To set up our experiments, our system is implemented ? Case 1: when training the CNN model using original
using Python 3.7 language and then using the programs of and testing with original images and all distorted images
the Anaconda Navigator and Spyder 4.1.4 environment on of three types of attacks are FGSM, Deep Fool, and
an Intel (R) 2.20 GHz Core (TM) i7-8750H CPU with 12.0 PGD to obtain a medium accuracy.
GB of RAM running Windows 10. So, the test is conducted
on five experiments, and each experiment includes, in the ? Case 2: when training the CNN model using original
train set and test set, a different number of images. Table 3 and all distorted images of three types of attacks are
illustrates the number of images in the train set and test set FGSM, Deep Fool, and PGD and testing with original
for each experiment. images, and the accuracy was high.

B. Results ? Case 3: when training the CNN model using original
and all distorted images of three types of attacks are
In all experiments, the CNN is trained with images FGSM, Deep Fool, and PGD and testing with original
generated using three types of attacks. The performance of and all distorted images and the accuracy is close to the
the proposed CNN is evaluated to recognize the difference accuracy of the original images.
between original images and adversarial images with the use
of a Rmsprop optimizer for perturbation optimization. In addition to our network, the database was trained
However, the first experiment is by training the CNN model using VGG-16 and VGG-19 network; however, the results
using the original image set. Therefore, the accuracy obtained were not very accurate compared to our model
obtained for the testing set is approximately 95%. which was of high accuracy in distinguishing between faces
in original and distorted images.
While the second experiment is by training the original
and distorted images using the FGSM, Deep Fool and, PGD

TABLE 3

NUMBER OF IMAGES IN EACH EXPERIMENT

Experiment No. original No. original No. image No. image (original No. image (original
image in train image in the (original & & distorted) in & distorted) in train
Clean distorted) in test train set
FGSM set test set set & test set
Deep fool set
PGD -
Merge 720 363 - - 5389
2154
720 363 1799 3590 2154
7543
720 363 718 1436

720 363 718 1436

720 363 2517 5026

TABLE 4

RECOGNITION FACE ACCURACY FOR SEVERAL OF SEVERAL SITUATIONS

Experiment Train set (Original) Train set (Original) Train set (Original Train & Test
& Test set (Original ) & Test set ( Original & & distorted) & (Original &
Clean Test (Original) Distorted)
FGSM Distorted)
Deep fool -
PGD 95% - - 93%
Merge 97%
- 54% 89% 95%
89%
- 70% 94%

- 94% 95%

- 60% 93%