Uploaded bykhalooei
333 views

Robustness of Deep Neural Networks

The document discusses the robustness of deep neural networks (DNN), focusing on their vulnerabilities to various types of attacks, including adversarial examples. It explores different attack models, defense strategies, and the complexities of maintaining DNN robustness amidst adversarial threats. Additionally, the document emphasizes the challenges associated with creating effective defenses against these adversarial attacks.

Embed presentation

Downloaded 15 times
Robustness of Deep Neural Networks Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 1 of 80
Robustness of Deep Neural Networks Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 2 of 80
• Decision-boundary crossing • Vulnerability of deep neural networks in different data • Different attacks and penetration ways of ML & DL approaches • Attack Toolboxes • Different defense strategies for DNN • Tips to stay safe in DNN Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Outlines 3 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 𝑑2 𝑑1 c 4 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 c 𝑑1 × 𝑑2 × 𝑐 0.975 , 0.2, 0.0023, 0.5, ….. , 0.5156 𝑑2 𝑑1 5 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data 6 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 7 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Life of the points … https://www.slideshare.net/khalooei/life-of-points-machine-learning-with-orange-flavor https://ceit.aut.ac.ir/~khalooei/presentations/ https://www.slideshare.net/khalooei/life-of-points-machine-learning-with-orange-flavor 8 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 9 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 10 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 11 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 12 of 84
Generative Model Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 TrainingData Training Generated Data Unseen new data 13 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 Distribution of the actual images 14 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 15 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 16 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 17 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 Distribution of the actual images 18 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Decision boundary crossing https://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html 19 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Decision boundary crossing http://www.cse.chalmers.se/~richajo/dit866/lectures/l3/Plotting%20decision%20boundaries.html Decision Tree Perceptron 20 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Vulnerability of Deep Neural Networks https://www.euclidean.com/deep-learning-and-value-investing 21 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation https://www.usenix.org/sites/default/files/conference/protected-files/enigma17_slides_papernot.pdf 22 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation 23 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation 24 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Machine learning … Adversarial Machine Learning, Author(s): Anthony D. Joseph, Blaine Nelson, Benjamin I. P. Rubinstein, J. D. Tygar Publisher: Cambridge University Press, Year: 2019 25 of 84
Fooling Google's image- recognition AI 1000x faster Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … ` December 20 '17 26 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … https://syncedreview.com/2019/04/24/now-you-see-me-now-you-dont-fooling-a-person-detector/ 27 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … https://towardsdatascience.com/evasion-attacks-on-machine-learning-or-adversarial-examples-12f2283e06a1 28 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir An interesting usage :: Google verification! 29 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … Audio Adversarial Examples: Targeted Attacks on Speech-to-Text Nicholas Carlini, David Wagner Deep Learning and Security Workshop, 2018. Best Paper 30 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … Generating Natural Language Adversarial Examples Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, Bo-Jhang Ho, Mani B. Srivastava, Kai-Wei Chang Published in EMNLP 2018 DOI:10.18653/v1/d18-1316 31 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Delving into the problem … 32 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation 33 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 34 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 35 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 36 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir https://www.slideshare.net/DavidKim486/universal-adversarial-perturbation 37 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir What is Adversarial Example? 𝑋 − ෠𝑋 𝑝 < 𝜀 𝑋 ෠𝑋 𝑋 ෠𝑋 39 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Examples of Adversarial Example https://www.kdnuggets.com/2018/10/adversarial-examples-explained.html ?! ?! 40 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir From WWW 41 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 42 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Perturbation's effect on class distributions 43 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir https://www.slideshare.net/DavidKim486/universal-adversarial-perturbation 44 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Examples from Overfitting • Adversarial Examples rooted in : • Overfitting • Excessive Linearity 45 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Examples from Excessive Linearity • Adversarial Examples rooted in : • Overfitting • Excessive Linearity 46 of 84
• The Attack Surface • The Adversarial Capabilities • The Adversarial Goals Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model 47 of 84
• The Attack Surface • Evasion Attack :: during the testing phase (* the most common type of attack!) • Poisoning Attack :: during the training time • Exploratory Attack :: during the testing phase (Given black box access to the model Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model try to gain as much knowledge as possible) http://fna.ir/a5g 48 of 84
• The Adversarial Capabilities • Training Phase Capabilities • Data Injection :: does not have any access to the training data as well as to the learning algorithm but has ability to • Data Modification :: does not have access to the learning algorithm but has full access to the training data • Logic Corruption :: meddle with the learning algorithm • Testing Phase Capabilities • White-Box Attacks :: an adversary has total knowledge about the model (f), algorithm (train), training data distribution (𝜇), • Black-Box Attacks :: no knowledge about the model and uses information about the settings or past inputs • Non-Adaptive Black-Box Attack :: only gets access to the target model’s training data distribution (μ) • Adaptive Black-Box Attack :: doesn’t have any information regarding the training process but can access the target model as an oracle • Strict Black-Box Attack :: may not contain the data distribution(μ) but has the ability to collect the input-output pairs(x,y) from the Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model augment a new data to the training set target classifier. However, he can not change the inputs to observe the changes in output like an adaptive attack procedure parameters (𝜃) of the fully trained model architecture http://fna.ir/a5g 50 of 84
• Adversarial Goals: • Confidence Reduction • The adversary tries to reduce the confidence of prediction for the target model • Misclassification • The adversary tries to alter the output classification of an input example to any class different from the original class. • Targeted Misclassification • The adversary tries to produce inputs that force the output of the classification model to be a specific target class • Source/Target Misclassification • The adversary attempts to force the output of classification for a specific input to be a particular target class Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model http://fna.ir/a5g 51 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model http://fna.ir/a5g 52 of 84
Adversarial example crafting procedures : 1. Direction sensitivity estimation • The adversary evaluate the sensitivity of a class change to each input feature • By identifying the direction in the data manifold around the example X 2. Perturbation selection • The adversary exploits the knowledge of sensitive information to select a perturbation • Selecting the perturbation 𝛿𝑋 3. Replace 𝑿 with 𝑿 + 𝜹𝑿 Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Example Generation Most DNN models make this formulation non-linear and non-convex, making it hard to find a closed-solution in most of the cases http://fna.ir/a5g 53 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Crafting adversarial examples: https://thomas-tanay.github.io/post--L2-regularization/http://fna.ir/a5g 55 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Defenses against Adversarial Examples Defenses http://www.rmmagazine.com/2016/06/01/the-3-lines-of-defense-for-good-risk-management/ 56 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review …. ?! ?! 57 of 84
❑Defense is hard! A theoretical model of the adversarial example crafting process is very difficult to construct. ▪ Non-linearity ▪ non-convex ▪ Complex optimization process ▪ … Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Advances in defense strategies ❑Most of the current defense strategies are ▪not adaptive to all types of adversarial attack ❑Implementation of such defense strategies ▪may incur performance overhead 58 of 84
• Removing perturbation with an autoencoder • Adding noise at test time • Ensembles • Confidence-reducing perturbation at test time • Dropout • Adding noise at train time • Various non-linear units • … Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Failed defenses http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture16.pdf 59 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial ☺ Timeline 60 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Defense Attack Concept 61 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 62 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 63 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Adversarial Training 64 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 65 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 7 𝑓 𝑥 + 𝜂 ≠ 𝑓 𝑥 𝑓𝑜𝑟 "𝑚𝑜𝑠𝑡" 𝑥~𝑋 𝜂 𝑝≤ 𝜉 ℙ 𝑥~𝑋 𝑓 𝑥 + 𝜂 ≠ 𝑓 𝑥 ≥ 1 − 𝛿 ∆𝜂 ← 𝑎𝑟𝑔 min 𝑟 𝑟 2 𝑠. 𝑡. 𝑓 𝑥𝑖 + 𝜂 + 𝑟 ≠ 𝑓 𝑥𝑖 𝒫 𝑝,ξ(η) = arg min 𝜂′ 𝜂 − 𝜂′ 2 𝑠. 𝑡. 𝜂′ 𝑝 ≤ 𝜉൯𝜂 ← 𝒫 𝑝,ξ(η + ∆𝜂𝑖 66 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 7 𝒫 𝑝,ξ(η) = arg min 𝜂′ 𝜂 − 𝜂′ 2 𝑠. 𝑡. 𝜂′ 𝑝 ≤ 𝜉 67 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 68 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Carlini 69 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Samangouei 70 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Tramèr 71 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Samangouei 72 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Stutz 73 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Trends on Adversarial learning http://fna.ir/a5d 643 0 783 1 979 8 1294 33 CVPR 2019 74 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Sabokrou, Khalooei, Adeli 75 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir List of toolboxes ToolBox Base Lib. Usability Updating Clever-Hans TensorFlow, Pytorch Semi Well (Oct 2019) Fool-Box TensorFlow, Keras, Pytorch, MXnet Easy Well (Oct 2019) IBM ART python Semi Well (Oct 2019) AdverTorch Python Semi Well (Dec 2019) … 77 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: CleverHans https://github.com/tensorflow/cleverhans http://www.cleverhans.io/ 78 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: CleverHans https://cleverhans.readthedocs.io/en/latest/ 79 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: FoolBox 80 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: FoolBox 81 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Tips to stay safe 82 of 84
• Decision-boundary crossing • Vulnerability of deep neural networks in different data • Different attacks and penetration ways of ML & DL approaches • Different defense strategies for DNN • Roadmap of the Adversarial Machine learning • Attack Toolboxes Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir A summary ☺ 83 of 84
Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 84 of 84

More Related Content

PDF
Support Vector Machines ( SVM )
byMohammad Junaid Khan
 
DOCX
Hand Written Character Recognition Using Neural Networks
byChiranjeevi Adi
 
PDF
Single Image Super Resolution Overview
byLEE HOSEONG
 
PPTX
Deep Learning Explained
byMelanie Swan
 
PDF
Deep learning - A Visual Introduction
byLukas Masuch
 
PDF
How much position information do convolutional neural networks encode? review...
byDongmin Choi
 
PPTX
CNN Tutorial
bySungjoon Choi
 
PDF
Autoencoders
byCloudxLab
 
Support Vector Machines ( SVM )
byMohammad Junaid Khan
 
Hand Written Character Recognition Using Neural Networks
byChiranjeevi Adi
 
Single Image Super Resolution Overview
byLEE HOSEONG
 
Deep Learning Explained
byMelanie Swan
 
Deep learning - A Visual Introduction
byLukas Masuch
 
How much position information do convolutional neural networks encode? review...
byDongmin Choi
 
CNN Tutorial
bySungjoon Choi
 
Autoencoders
byCloudxLab
 

What's hot

PPT
Multi-Layer Perceptrons
byESCOM
 
PPTX
Using SHAP to Understand Black Box Models
byJonathan Bechtel
 
PDF
Deep Belief Networks
byHasan H Topcu
 
PPTX
Machine Learning Algorithms | Machine Learning Tutorial | Data Science Algori...
bySimplilearn
 
PPTX
Multilayer perceptron
byomaraldabash
 
PPTX
Introduction to CNN Models: DenseNet & MobileNet
byKrishnakoumarC
 
PDF
Neural Networks: Multilayer Perceptron
byMostafa G. M. Mostafa
 
PDF
Convolutional neural network
byYan Xu
 
PPTX
Support Vector Machine - How Support Vector Machine works | SVM in Machine Le...
bySimplilearn
 
PDF
Intro To Convolutional Neural Networks
byMark Scully
 
PPTX
Data Preprocessing and Feature Engineering
byShehzadBachani1
 
PPTX
Autoencoders in Deep Learning
bymilad abbasi
 
PPTX
Introduction to Deep Learning
byOswald Campesato
 
PPTX
Machine Learning Project
byAbhishek Singh
 
PDF
Deep learning
byMohamed Loey
 
PDF
Image segmentation with deep learning
byAntonio Rueda-Toicen
 
PPT
K mean-clustering algorithm
byparry prabhu
 
PDF
An Introduction to Supervised Machine Learning and Pattern Classification: Th...
bySebastian Raschka
 
PDF
Bayesian networks in AI
byByoung-Hee Kim
 
PPTX
Darknet yolo
byBang Tsui Liou
 
Multi-Layer Perceptrons
byESCOM
 
Using SHAP to Understand Black Box Models
byJonathan Bechtel
 
Deep Belief Networks
byHasan H Topcu
 
Machine Learning Algorithms | Machine Learning Tutorial | Data Science Algori...
bySimplilearn
 
Multilayer perceptron
byomaraldabash
 
Introduction to CNN Models: DenseNet & MobileNet
byKrishnakoumarC
 
Neural Networks: Multilayer Perceptron
byMostafa G. M. Mostafa
 
Convolutional neural network
byYan Xu
 
Support Vector Machine - How Support Vector Machine works | SVM in Machine Le...
bySimplilearn
 
Intro To Convolutional Neural Networks
byMark Scully
 
Data Preprocessing and Feature Engineering
byShehzadBachani1
 
Autoencoders in Deep Learning
bymilad abbasi
 
Introduction to Deep Learning
byOswald Campesato
 
Machine Learning Project
byAbhishek Singh
 
Deep learning
byMohamed Loey
 
Image segmentation with deep learning
byAntonio Rueda-Toicen
 
K mean-clustering algorithm
byparry prabhu
 
An Introduction to Supervised Machine Learning and Pattern Classification: Th...
bySebastian Raschka
 
Bayesian networks in AI
byByoung-Hee Kim
 
Darknet yolo
byBang Tsui Liou
 

Similar to Robustness of Deep Neural Networks

PDF
Robustness of compressed CNNs
byKaushalya Madhawa
 
PDF
DEF CON 24 - Clarence Chio - machine duping 101
byFelipe Prado
 
PDF
Generational Adversarial Neural Networks - Essential Reference
byGokul Alex
 
PDF
Robustness of Deep Neural Networks on White-box Attacks and Defense Strategie...
byAkhileshPandey104
 
PDF
Machine Duping 101: Pwning Deep Learning Systems
byClarence Chio
 
PPTX
adversarial robustness lecture
byMuhammadAhmedShah2
 
PDF
introduction to deeplearning
byEyad Alshami
 
PPT
deeplearning
byhuda2018
 
PDF
Detecting adversarials examples attacks to deep neural networks
byFabio Carrara
 
PPTX
slides_security_and_privacy_in_machine_learning.pptx
byssuserabf73f
 
PDF
005 20151130 adversary_networks
byHa Phuong
 
PDF
Deep learning: Cutting through the Myths and Hype
bySiby Jose Plathottam
 
PDF
Adversarial examples in deep learning (Gregory Chatel)
byMeetupDataScienceRoma
 
PPTX
What Is Deep Learning? | Introduction to Deep Learning | Deep Learning Tutori...
bySimplilearn
 
PDF
Deep Learning - Overview of my work II
byMohamed Loey
 
PDF
BOOSTING ADVERSARIAL ATTACKS WITH MOMENTUM - Tianyu Pang and Chao Du, THU - D...
byGeekPwn Keen
 
PPTX
Deep Learning Tutorial
byAmr Rashed
 
PPTX
Deep learning tutorial 9/2019
byAmr Rashed
 
PPTX
Introduction to deep learning
bydoppenhe
 
PDF
Deep Learning Frameworks slides
bySheamus McGovern
 
Robustness of compressed CNNs
byKaushalya Madhawa
 
DEF CON 24 - Clarence Chio - machine duping 101
byFelipe Prado
 
Generational Adversarial Neural Networks - Essential Reference
byGokul Alex
 
Robustness of Deep Neural Networks on White-box Attacks and Defense Strategie...
byAkhileshPandey104
 
Machine Duping 101: Pwning Deep Learning Systems
byClarence Chio
 
adversarial robustness lecture
byMuhammadAhmedShah2
 
introduction to deeplearning
byEyad Alshami
 
deeplearning
byhuda2018
 
Detecting adversarials examples attacks to deep neural networks
byFabio Carrara
 
slides_security_and_privacy_in_machine_learning.pptx
byssuserabf73f
 
005 20151130 adversary_networks
byHa Phuong
 
Deep learning: Cutting through the Myths and Hype
bySiby Jose Plathottam
 
Adversarial examples in deep learning (Gregory Chatel)
byMeetupDataScienceRoma
 
What Is Deep Learning? | Introduction to Deep Learning | Deep Learning Tutori...
bySimplilearn
 
Deep Learning - Overview of my work II
byMohamed Loey
 
BOOSTING ADVERSARIAL ATTACKS WITH MOMENTUM - Tianyu Pang and Chao Du, THU - D...
byGeekPwn Keen
 
Deep Learning Tutorial
byAmr Rashed
 
Deep learning tutorial 9/2019
byAmr Rashed
 
Introduction to deep learning
bydoppenhe
 
Deep Learning Frameworks slides
bySheamus McGovern
 

More from khalooei

PDF
Robustness of Deep Neural Networks | Adversarial attacks and defenses
bykhalooei
 
PDF
Life of Points (Machine learning with Orange flavor)
bykhalooei
 
PDF
Generative Adversarial Network
bykhalooei
 
PDF
Generative Adversarial Networks - (Applications)
bykhalooei
 
PDF
Generative Adversarial Networks - (Introduction)
bykhalooei
 
PDF
تحلیل با رویکرد یادگیری ژرف بر بستر کلان‌داده (2)
bykhalooei
 
PDF
تحلیل با رویکرد یادگیری ژرف بر بستر کلان‌داده
bykhalooei
 
Robustness of Deep Neural Networks | Adversarial attacks and defenses
bykhalooei
 
Life of Points (Machine learning with Orange flavor)
bykhalooei
 
Generative Adversarial Network
bykhalooei
 
Generative Adversarial Networks - (Applications)
bykhalooei
 
Generative Adversarial Networks - (Introduction)
bykhalooei
 
تحلیل با رویکرد یادگیری ژرف بر بستر کلان‌داده (2)
bykhalooei
 
تحلیل با رویکرد یادگیری ژرف بر بستر کلان‌داده
bykhalooei
 

Recently uploaded

PPTX
Chapter 1: Introduction to Strategic Management.pptx
byFJ M
 
PDF
Pratishta Educational Society., Courses & Opportunities
byGanapathiVankudoth
 
PPTX
Lesson_1 Acceleration.pptx_Science 8-4th quarter
byAnnabelAlarcon
 
PDF
Unit3 Phases and process of curriculum development.pdf
byziaoria016
 
PPTX
Greengnorance Toolkit Module 4 Active Mobility and Transport
byKarl Donert
 
PDF
Nursing care plan for Vomiting /B.Sc nsg
byDr. Sudhadevi Sadanandan
 
PDF
Judgement Regarding Land Acquisition Act not Applicable to Encroachers.pdf
byssuser9d8e4e
 
PPTX
Math 8 Quarter 4 Week 4-SECONDARY DATA.pptx
byshahanieabbat3
 
PDF
Tetracycline Class of Antibiotics: SAR, MOA and Uses
bymominzarinamdnajeeb
 
PDF
Workshop 29 Crystal Review All Students by YogiGoddess
by©LDMMIA, ©Reiki Yoga
 
PDF
Chapter 05 Drugs Acting on the Central Nervous System: Anti-Depressant Drugs
bySandeshSul
 
PDF
How "Raiders of the Lost Ark" and "Ordinary People" Employ Non-Verbal Acting
by6x5csns4pn
 
PPTX
Greengnorance Toolkit Module 5 Waste Management
byKarl Donert
 
PPTX
Plato's Insight model teaching and learning.pptx
byNikhitaDS
 
PPTX
Problem and Solution (PowerPoint Game Template)
byacpaulite
 
PDF
RAJAT ARORA SIR PHYSICAL EDUCATION NOTES ALL CHAPTERS .pdf
bysumitkannoujiya74
 
PDF
Unit Plan and Unit Test-pdf-Dr. Rajashekhar Shirvalkar, Principal, SMRS B.Ed ...
byRajashekhar Shirvalkar
 
PDF
Types of Foot & Foot Modifications in Birds
byDr. Ramzan Virani
 
PPTX
ECP Demo ppt.pptx Visualizing and Identifying solid figures using concrete an...
byireneodechavez1
 
PPTX
How to Track & Manage My Time Section in Odoo 18 Time Off
byCeline George
 
Chapter 1: Introduction to Strategic Management.pptx
byFJ M
 
Pratishta Educational Society., Courses & Opportunities
byGanapathiVankudoth
 
Lesson_1 Acceleration.pptx_Science 8-4th quarter
byAnnabelAlarcon
 
Unit3 Phases and process of curriculum development.pdf
byziaoria016
 
Greengnorance Toolkit Module 4 Active Mobility and Transport
byKarl Donert
 
Nursing care plan for Vomiting /B.Sc nsg
byDr. Sudhadevi Sadanandan
 
Judgement Regarding Land Acquisition Act not Applicable to Encroachers.pdf
byssuser9d8e4e
 
Math 8 Quarter 4 Week 4-SECONDARY DATA.pptx
byshahanieabbat3
 
Tetracycline Class of Antibiotics: SAR, MOA and Uses
bymominzarinamdnajeeb
 
Workshop 29 Crystal Review All Students by YogiGoddess
by©LDMMIA, ©Reiki Yoga
 
Chapter 05 Drugs Acting on the Central Nervous System: Anti-Depressant Drugs
bySandeshSul
 
How "Raiders of the Lost Ark" and "Ordinary People" Employ Non-Verbal Acting
by6x5csns4pn
 
Greengnorance Toolkit Module 5 Waste Management
byKarl Donert
 
Plato's Insight model teaching and learning.pptx
byNikhitaDS
 
Problem and Solution (PowerPoint Game Template)
byacpaulite
 
RAJAT ARORA SIR PHYSICAL EDUCATION NOTES ALL CHAPTERS .pdf
bysumitkannoujiya74
 
Unit Plan and Unit Test-pdf-Dr. Rajashekhar Shirvalkar, Principal, SMRS B.Ed ...
byRajashekhar Shirvalkar
 
Types of Foot & Foot Modifications in Birds
byDr. Ramzan Virani
 
ECP Demo ppt.pptx Visualizing and Identifying solid figures using concrete an...
byireneodechavez1
 
How to Track & Manage My Time Section in Odoo 18 Time Off
byCeline George
 

Robustness of Deep Neural Networks

  • 1.
    Robustness of DeepNeural Networks Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 1 of 80
  • 2.
    Robustness of DeepNeural Networks Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir 2 of 80
  • 3.
    • Decision-boundary crossing •Vulnerability of deep neural networks in different data • Different attacks and penetration ways of ML & DL approaches • Attack Toolboxes • Different defense strategies for DNN • Tips to stay safe in DNN Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Outlines 3 of 84
  • 4.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 𝑑2 𝑑1 c 4 of 84
  • 5.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 c 𝑑1 × 𝑑2 × 𝑐 0.975 , 0.2, 0.0023, 0.5, ….. , 0.5156 𝑑2 𝑑1 5 of 84
  • 6.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data 6 of 84
  • 7.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Data https://www.slideshare.net/BrianKim244/dcgan-77452250 7 of 84
  • 8.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Life of the points … https://www.slideshare.net/khalooei/life-of-points-machine-learning-with-orange-flavor https://ceit.aut.ac.ir/~khalooei/presentations/ https://www.slideshare.net/khalooei/life-of-points-machine-learning-with-orange-flavor 8 of 84
  • 9.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 9 of 84
  • 10.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 10 of 84
  • 11.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 11 of 84
  • 12.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Discriminative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 12 of 84
  • 13.
    Generative Model Robustness ofDeep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 TrainingData Training Generated Data Unseen new data 13 of 84
  • 14.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 Distribution of the actual images 14 of 84
  • 15.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 15 of 84
  • 16.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 16 of 84
  • 17.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 17 of 84
  • 18.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review:: Generative Model https://www.slideshare.net/BrianKim244/dcgan-77452250 Distribution of the actual images 18 of 84
  • 19.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Decision boundary crossing https://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html 19 of 84
  • 20.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Decision boundary crossing http://www.cse.chalmers.se/~richajo/dit866/lectures/l3/Plotting%20decision%20boundaries.html Decision Tree Perceptron 20 of 84
  • 21.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Vulnerability of Deep Neural Networks https://www.euclidean.com/deep-learning-and-value-investing 21 of 84
  • 22.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation https://www.usenix.org/sites/default/files/conference/protected-files/enigma17_slides_papernot.pdf 22 of 84
  • 23.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation 23 of 84
  • 24.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir AI and the top innovation 24 of 84
  • 25.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Machine learning … Adversarial Machine Learning, Author(s): Anthony D. Joseph, Blaine Nelson, Benjamin I. P. Rubinstein, J. D. Tygar Publisher: Cambridge University Press, Year: 2019 25 of 84
  • 26.
    Fooling Google's image- recognitionAI 1000x faster Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … ` December 20 '17 26 of 84
  • 27.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … https://syncedreview.com/2019/04/24/now-you-see-me-now-you-dont-fooling-a-person-detector/ 27 of 84
  • 28.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … https://towardsdatascience.com/evasion-attacks-on-machine-learning-or-adversarial-examples-12f2283e06a1 28 of 84
  • 29.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir An interesting usage :: Google verification! 29 of 84
  • 30.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … Audio Adversarial Examples: Targeted Attacks on Speech-to-Text Nicholas Carlini, David Wagner Deep Learning and Security Workshop, 2018. Best Paper 30 of 84
  • 31.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir On the line of recent researches … Generating Natural Language Adversarial Examples Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, Bo-Jhang Ho, Mani B. Srivastava, Kai-Wei Chang Published in EMNLP 2018 DOI:10.18653/v1/d18-1316 31 of 84
  • 32.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Delving into the problem … 32 of 84
  • 33.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation 33 of 84
  • 34.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 34 of 84
  • 35.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 35 of 84
  • 36.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Motivation Gradient-descent 36 of 84
  • 37.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir https://www.slideshare.net/DavidKim486/universal-adversarial-perturbation 37 of 84
  • 38.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir What is Adversarial Example? 𝑋 − ෠𝑋 𝑝 < 𝜀 𝑋 ෠𝑋 𝑋 ෠𝑋 39 of 84
  • 39.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Examples of Adversarial Example https://www.kdnuggets.com/2018/10/adversarial-examples-explained.html ?! ?! 40 of 84
  • 40.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir From WWW 41 of 84
  • 41.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir 42 of 84
  • 42.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Perturbation's effect on class distributions 43 of 84
  • 43.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir https://www.slideshare.net/DavidKim486/universal-adversarial-perturbation 44 of 84
  • 44.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Examples from Overfitting • Adversarial Examples rooted in : • Overfitting • Excessive Linearity 45 of 84
  • 45.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Examples from Excessive Linearity • Adversarial Examples rooted in : • Overfitting • Excessive Linearity 46 of 84
  • 46.
    • The AttackSurface • The Adversarial Capabilities • The Adversarial Goals Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model 47 of 84
  • 47.
    • The AttackSurface • Evasion Attack :: during the testing phase (* the most common type of attack!) • Poisoning Attack :: during the training time • Exploratory Attack :: during the testing phase (Given black box access to the model Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model try to gain as much knowledge as possible) http://fna.ir/a5g 48 of 84
  • 48.
    • The AdversarialCapabilities • Training Phase Capabilities • Data Injection :: does not have any access to the training data as well as to the learning algorithm but has ability to • Data Modification :: does not have access to the learning algorithm but has full access to the training data • Logic Corruption :: meddle with the learning algorithm • Testing Phase Capabilities • White-Box Attacks :: an adversary has total knowledge about the model (f), algorithm (train), training data distribution (𝜇), • Black-Box Attacks :: no knowledge about the model and uses information about the settings or past inputs • Non-Adaptive Black-Box Attack :: only gets access to the target model’s training data distribution (μ) • Adaptive Black-Box Attack :: doesn’t have any information regarding the training process but can access the target model as an oracle • Strict Black-Box Attack :: may not contain the data distribution(μ) but has the ability to collect the input-output pairs(x,y) from the Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model augment a new data to the training set target classifier. However, he can not change the inputs to observe the changes in output like an adaptive attack procedure parameters (𝜃) of the fully trained model architecture http://fna.ir/a5g 50 of 84
  • 49.
    • Adversarial Goals: •Confidence Reduction • The adversary tries to reduce the confidence of prediction for the target model • Misclassification • The adversary tries to alter the output classification of an input example to any class different from the original class. • Targeted Misclassification • The adversary tries to produce inputs that force the output of the classification model to be a specific target class • Source/Target Misclassification • The adversary attempts to force the output of classification for a specific input to be a particular target class Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model http://fna.ir/a5g 51 of 84
  • 50.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Threat Model http://fna.ir/a5g 52 of 84
  • 51.
    Adversarial example craftingprocedures : 1. Direction sensitivity estimation • The adversary evaluate the sensitivity of a class change to each input feature • By identifying the direction in the data manifold around the example X 2. Perturbation selection • The adversary exploits the knowledge of sensitive information to select a perturbation • Selecting the perturbation 𝛿𝑋 3. Replace 𝑿 with 𝑿 + 𝜹𝑿 Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial Example Generation Most DNN models make this formulation non-linear and non-convex, making it hard to find a closed-solution in most of the cases http://fna.ir/a5g 53 of 84
  • 52.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Crafting adversarial examples: https://thomas-tanay.github.io/post--L2-regularization/http://fna.ir/a5g 55 of 84
  • 53.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Defenses against Adversarial Examples Defenses http://www.rmmagazine.com/2016/06/01/the-3-lines-of-defense-for-good-risk-management/ 56 of 84
  • 54.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Review …. ?! ?! 57 of 84
  • 55.
    ❑Defense is hard! Atheoretical model of the adversarial example crafting process is very difficult to construct. ▪ Non-linearity ▪ non-convex ▪ Complex optimization process ▪ … Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Advances in defense strategies ❑Most of the current defense strategies are ▪not adaptive to all types of adversarial attack ❑Implementation of such defense strategies ▪may incur performance overhead 58 of 84
  • 56.
    • Removing perturbationwith an autoencoder • Adding noise at test time • Ensembles • Confidence-reducing perturbation at test time • Dropout • Adding noise at train time • Various non-linear units • … Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir Failed defenses http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture16.pdf 59 of 84
  • 57.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial ☺ Timeline 60 of 84
  • 58.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Defense Attack Concept 61 of 84
  • 59.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 62 of 84
  • 60.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 63 of 84
  • 61.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Adversarial Training 64 of 84
  • 62.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 65 of 84
  • 63.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 7 𝑓 𝑥 + 𝜂 ≠ 𝑓 𝑥 𝑓𝑜𝑟 "𝑚𝑜𝑠𝑡" 𝑥~𝑋 𝜂 𝑝≤ 𝜉 ℙ 𝑥~𝑋 𝑓 𝑥 + 𝜂 ≠ 𝑓 𝑥 ≥ 1 − 𝛿 ∆𝜂 ← 𝑎𝑟𝑔 min 𝑟 𝑟 2 𝑠. 𝑡. 𝑓 𝑥𝑖 + 𝜂 + 𝑟 ≠ 𝑓 𝑥𝑖 𝒫 𝑝,ξ(η) = arg min 𝜂′ 𝜂 − 𝜂′ 2 𝑠. 𝑡. 𝜂′ 𝑝 ≤ 𝜉൯𝜂 ← 𝒫 𝑝,ξ(η + ∆𝜂𝑖 66 of 84
  • 64.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 7 𝒫 𝑝,ξ(η) = arg min 𝜂′ 𝜂 − 𝜂′ 2 𝑠. 𝑡. 𝜂′ 𝑝 ≤ 𝜉 67 of 84
  • 65.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … 68 of 84
  • 66.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Carlini 69 of 84
  • 67.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Samangouei 70 of 84
  • 68.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Tramèr 71 of 84
  • 69.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Samangouei 72 of 84
  • 70.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Stutz 73 of 84
  • 71.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Trends on Adversarial learning http://fna.ir/a5d 643 0 783 1 979 8 1294 33 CVPR 2019 74 of 84
  • 72.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Adversarial timeline … Sabokrou, Khalooei, Adeli 75 of 84
  • 73.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir List of toolboxes ToolBox Base Lib. Usability Updating Clever-Hans TensorFlow, Pytorch Semi Well (Oct 2019) Fool-Box TensorFlow, Keras, Pytorch, MXnet Easy Well (Oct 2019) IBM ART python Semi Well (Oct 2019) AdverTorch Python Semi Well (Dec 2019) … 77 of 84
  • 74.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: CleverHans https://github.com/tensorflow/cleverhans http://www.cleverhans.io/ 78 of 84
  • 75.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: CleverHans https://cleverhans.readthedocs.io/en/latest/ 79 of 84
  • 76.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: FoolBox 80 of 84
  • 77.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir ToolBoxes :: FoolBox 81 of 84
  • 78.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir Tips to stay safe 82 of 84
  • 79.
    • Decision-boundary crossing •Vulnerability of deep neural networks in different data • Different attacks and penetration ways of ML & DL approaches • Different defense strategies for DNN • Roadmap of the Adversarial Machine learning • Attack Toolboxes Robustness of Deep Neural Networks Mohammad Khalooei | khalooei@aut.ac.ir A summary ☺ 83 of 84
  • 80.
    Robustness of DeepNeural Networks Mohammad Khalooei | khalooei@aut.ac.ir 84 of 84