Robustness of Deep learning mode ls.pptx

Thanh Nguyen Cong - Robustness of Deep learning models

1. Robustness of Deep learning models Author: Nguyễn Công Thạnh Email: [email protected] Phone: [phone]

2. Agenda 1. Introduction 2. Related Works 3. Hybrid dataset 3.1. Introduction of the Hybrid dataset 3.2. Evaluation measures 3.3. The description of Hybrid dataset 3.4. Comparison 4. Experiments and Results 5. Conclusions 2

3. 1. Introduction Iris dataset 3 MNIST dataset CIFAR-10 dataset Deep Learning models + = Classification models & High accuracy Adversarial attacks !!!

4. 1. Introduction 4 Why choose CIFAR-10 to generate Hybrid CIFAR-10? A widely recognized benchmark for image classification. Represents everyday objects like airplanes, cars, birds, and cats. Enables faster training and easier experimentation compared to CIFAR-100. CIFAR-10 dataset

5. 2. Related Works Attack Methods : FGSM, DeepFool , AdvGAN Generate adversarial images Merge all images and relabel to get new dataset Adversarial training to enhance the Robustness of Deep learning models Related works : Zhang et al. generated an adversarial images from MNIST and CIFAR-10 using LSGAN. Jordan et al. introduced the CIFAKE dataset, using diffusion algorithms to create CIFAR-10 like images. Automated generation of adaptive perturbed images based on GAN for motivated adversaries on deep learning models: Automated generation of adaptive perturbed images based on GAN DOI:  10.1145/3628797.3628923 5 Adversarial example + = cat’s image perturbation adversarial image It’s still a cat! It’s a dog! human’s eye DL models !!!

6. 3. Hybrid dataset Introduction of the Hybrid CIFAR-10 dataset: 6 Some images in the Hybrid CIFAR-10 dataset (left images are origin images, right images is adversarial images generated by GAN method)

7. 3. Hybrid dataset Evaluation measures: Structural Similarity Index Measurement ( SSIM ): Peak Signal to Noise Ratio ( PSNR ): where x is value of pixels in the image, I is the original image, I′ is the converted image, I and I′ are 2-dimensional images of size m×n . 7

8. 3. Hybrid CIFAR-10 dataset The description of Hybrid CIFAR-10 dataset: 8 Data collection Dataset detail 158, 498 color images of size 32 × 32 across 10 classes. Structure:  image perturbation adversarial image GAN model  + Hybrid CIFAR-10 collect collect

9. 3. Hybrid CIFAR-10 dataset The description of Hybrid CIFAR-10 dataset: Evaluate the dataset: Statistics of SSIM and PSNR values according to class name Statistics of SSIM and PSNR measures according to k 9

10. 3. Hybrid CIFAR-10 dataset The description of Hybrid CIFAR-10 dataset: Usage: Useful for evaluating object classification models’ performance and reliability H elps test accuracy and robustness S erve as a supplementary training resource to enhance model robustness 10

11. 3. Hybrid CIFAR-10 dataset Comparison 11 Comparison of our proposed dataset with CIFAR-10 dataset Comparison of SSIM and PSNR measures of proposed dataset with previous studies (according to Zhang et al. )

12. 4. Experiments and Results Experiments Environment : Anaconda Python 3.11.3 PyTorch 1.8 CUDA 10.1 Intel Core i9-9900 CPU, 64 GB RAM NVIDIA GeForce RTX 2080 GPU, 8 GB VRAM Advanced image recognition models : ResNet-56 MobileNetV2 VGG19_bn ShuffleNetV2 RepVGG_a2 12 Accuracy (Acc): Precision (P): Recall (R):

13. 4. Experiments and Results Results 13 Accuracy (%) of the models when evaluated on Real images (Real) and Adversarial images (AEs) before and after Adversarial training ( AE training ). Results (%) of training models on Hybrid CIFAR-10 dataset.

14. 5. Conclusions We propose a dataset by combining original CIFAR-10 images with adversarial images generated by GAN method at different perturbation coefficient named Hybrid CIFAR-10 . Experimental results demonstrate its suitability for evaluating machine learning models and for resistance training to enhance model robustness and accuracy. In the future, we aim to develop larger , more diverse datasets using similar methods to enhance data variety, provide a continuous data source for training and improving the robustness of various deep learning models. 14