SlideShare uma empresa Scribd logo
1 de 26
Baixar para ler offline
Adversarial Attacks and
Defenses in Deep Learning
Kui Ren, Tiahnhang Zheng, Zhan Qin, Xue Liu (2020) Engineering
On April 23, 2013, Syrian hackers compromised the Associated Press
Twitter feed and tweeted, “Breaking: Two Explosions in the White House
and Barack Obama is injured”.
In response to the tweet, the Dow Jones Industrial Average dropped by
$136 billion dollars (although this drop was reversed 3 minutes later).
What are adversarial attacks and why should you
care?
● Any attempt to fool a deep learning model with deceptive input
● Especially researched in image recognition, but can also be applied to audio, text or tabular data
● When building models, we mostly focus on classiïŹcation effectiveness/ minimizing error. Relatively
little work on model security and robustness.
● Imperceptible amounts of non-random noise can fool neural networks!
● Some of these attacks are 100% effective in fooling normal neural networks!
Gong & Poellabauer, 2018. Comp Sci & Engineering
What I’ll talk about
● Threat-models
● Some background terminology
● Notable adversarial models
● Notable adversarial defences
● Trends and remaining challenges
● Code
Level of threat
● White-box: full knowledge of model architecture and parameters
● Gray-box: knowledge limited to features, model type
● Black-box: no/minimal knowledge of the model, can only use output
All non-adversarially trained models are susceptible, even to black box models
Adversarially trained models are still susceptible to white box models
Background
Adversarial loss: J(Ξ, x, y), Ξ = model weights
An adversarial sample x’ : D(x,x’) < η (predeïŹned distance constraint, perturbation)
● Idea: ïŹnd the minimum difference or perturbation f(x’) ≠ y’
Adversarial samples should be indistinguishable from
benign samples
Distance metrics:
● L₂ distance: What is the squared difference between adversarial and benign image?
● L∞ distance: Maximum element-wise difference between adversarial and benign image
(for each pixel, take the absolute value difference between X and Z, and return the
largest such distance found over all pixels)
Notable adversarial models
Limited-memory BFGS
Grid search / line search to ïŹnd optimal hyperparameter
Carlini and Wagner (C&W) attack
Set of optimization-based attacks that generate L₀, L₂ and L∞ norm measured adversarial samples, with some
restrictions (kappa) to make sure a valid image is produced
100% attack success on ‘normal’ neural networks trained on MNIST, CIFAR-10, ImageNet
Compromised defensive models
Notable adversarial models
DeepFool
“Iterative linearization of the
classiïŹer to generate minimal
perturbations that are sufïŹcient
to change classiïŹcation labels”
Computes perturbations more
reliably
Moosavi-Dezfooli et al., https://arxiv.org/pdf/1511.04599.pdf
Notable adversarial models
Universal adversarial attack
● Is there a universal perturbation that will work on most
samples?
● L-BFGS- based
● Effective in attacking NN like CaffeNet, GoogleNet, VGG and
Resnet
● Fooling rate > 53%
Adversarial patch
Only certain areas are perturbed
Text & Audio Models
1% audio perturbation can change 50 words in
text transcription!
Attacks are robust to MP3 compression, but get
lost when played over speakers
https://nicholas.carlini.com/code/audio_adversarial_examples/
Strategies for text attacks generally include
deleting, inserting, and modifying
characters/words
Adversarial defenses fall into 5 categories
1. Training on adversarial samples
2. Randomization
3. Adding noise
4. Removing noise
5. Mathematically provable defenses
Defang: Randomize input or features
● Randomly padding and resizing input; image transformations with randomness
● Add random noise layer before each convolutional layer in training and test sets (RSE)
● Random feature pruning at each layer
Detect: Denoise the input or features
● Conventional input rectiïŹcation
○ ‘Squeeze’ image → if output is very different from input, then likely adversarial
● GAN-based
○ Use GAN to learn benign data distribution
○ Generate a benign projection for the adversarial sample
● Autoencoder-based
○ Detector & reformer
○ Use an autocoder to compress input and learn manifold of benign samples
○ Detector compares each sample to learnt manifold
○ Reformer rectiïŹes adversarial samples
Detect: Denoise the input or features
● High-level representation guided denoiser (HGD)
○ Trains a denoising u-net using a feature-level loss function to minimize feature
differences between benign and adversarial samples
○ Won ïŹrst place in black-box defenses, 2017
○ Even so, certain (white-box) attacks can reduce effectiveness to 0%
Provable (certiïŹcated) defenses
● Defenses that have theoretical backing to have a certain accuracy against attacks
● Range of defenses include KNN and Bayesian-based defenses
● Consistency-based defenses:
○ Perturbations also affect the area around them
○ > 90 detection rate
● Very computationally intensive
Trends in adversarial research
● Design stronger attacks to probe for weaknesses
● Real-world attack capabilities
● CertiïŹcated defenses - but currently not scalable
“A problem is that an attack can only target one category of defenses, but defenses are required to 
 be effective
against all possible attack methods”
● Analyzing model robustness - mostly done on KKN and linear classiïŹers
Unresolved challenges
● Causality
● Does a general robust decision boundary exist that could be learnt by (certain) neural
networks?
● Effectiveness vs efïŹciency
○ Adversarial training is effective, but requires a lot of data and compute
○ Randomization and denoising strategies very efïŹcient, but not as effective as claimed
Discussion
In what other ways are models not robust?
Is model robustness/ security applicable to what you do / to our students?
Model fairness has been a hot topic lately, but robustness/ security seems to lag behind - what do you
think needs to change for adversarial training to be widely implemented?
What are your thoughts on the paper in general?
Try it yourself
Benchmark machine learning systems' vulnerability to adversarial examples:
https://github.com/cleverhans-lab/cleverhans
Blog: cleverhans.io

Mais conteĂșdo relacionado

Mais procurados

Malware Dectection Using Machine learning
Malware Dectection Using Machine learningMalware Dectection Using Machine learning
Malware Dectection Using Machine learningShubham Dubey
 
Basic Generative Adversarial Networks
Basic Generative Adversarial NetworksBasic Generative Adversarial Networks
Basic Generative Adversarial NetworksDong Heon Cho
 
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptx
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptxLecture_1_Introduction_to_Adversarial_Machine_Learning.pptx
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptxDevRaj646424
 
Privacy preserving machine learning
Privacy preserving machine learningPrivacy preserving machine learning
Privacy preserving machine learningMichaƂ KuĆșba
 
Robustness of Deep Neural Networks
Robustness of Deep Neural NetworksRobustness of Deep Neural Networks
Robustness of Deep Neural Networkskhalooei
 
Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Appsilon Data Science
 
Adversarial machine learning
Adversarial machine learningAdversarial machine learning
Adversarial machine learningRama Chetan
 
Generative Adversarial Networks and Their Applications
Generative Adversarial Networks and Their ApplicationsGenerative Adversarial Networks and Their Applications
Generative Adversarial Networks and Their ApplicationsArtifacia
 
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
 A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs) A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)Thomas da Silva Paula
 
Deep Learning for Computer Vision: Generative models and adversarial training...
Deep Learning for Computer Vision: Generative models and adversarial training...Deep Learning for Computer Vision: Generative models and adversarial training...
Deep Learning for Computer Vision: Generative models and adversarial training...Universitat PolitĂšcnica de Catalunya
 
Explainable Machine Learning (Explainable ML)
Explainable Machine Learning (Explainable ML)Explainable Machine Learning (Explainable ML)
Explainable Machine Learning (Explainable ML)Hayim Makabee
 
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAIGenerative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAIWithTheBest
 
Transfer learning-presentation
Transfer learning-presentationTransfer learning-presentation
Transfer learning-presentationBushra Jbawi
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial NetworksMustafa Yagmur
 
GAN - Theory and Applications
GAN - Theory and ApplicationsGAN - Theory and Applications
GAN - Theory and ApplicationsEmanuele Ghelfi
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networksë‚šìŁŒ êč€
 
Machine learning in Cyber Security
Machine learning in Cyber SecurityMachine learning in Cyber Security
Machine learning in Cyber SecurityRajathV2
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networksYunjey Choi
 
Unsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANUnsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANShyam Krishna Khadka
 

Mais procurados (20)

Malware Dectection Using Machine learning
Malware Dectection Using Machine learningMalware Dectection Using Machine learning
Malware Dectection Using Machine learning
 
Basic Generative Adversarial Networks
Basic Generative Adversarial NetworksBasic Generative Adversarial Networks
Basic Generative Adversarial Networks
 
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptx
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptxLecture_1_Introduction_to_Adversarial_Machine_Learning.pptx
Lecture_1_Introduction_to_Adversarial_Machine_Learning.pptx
 
Privacy preserving machine learning
Privacy preserving machine learningPrivacy preserving machine learning
Privacy preserving machine learning
 
Robustness of Deep Neural Networks
Robustness of Deep Neural NetworksRobustness of Deep Neural Networks
Robustness of Deep Neural Networks
 
Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)Introduction to Generative Adversarial Networks (GANs)
Introduction to Generative Adversarial Networks (GANs)
 
Adversarial machine learning
Adversarial machine learningAdversarial machine learning
Adversarial machine learning
 
Generative Adversarial Networks and Their Applications
Generative Adversarial Networks and Their ApplicationsGenerative Adversarial Networks and Their Applications
Generative Adversarial Networks and Their Applications
 
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
 A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs) A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
 
Deep Learning for Computer Vision: Generative models and adversarial training...
Deep Learning for Computer Vision: Generative models and adversarial training...Deep Learning for Computer Vision: Generative models and adversarial training...
Deep Learning for Computer Vision: Generative models and adversarial training...
 
Generative adversarial text to image synthesis
Generative adversarial text to image synthesisGenerative adversarial text to image synthesis
Generative adversarial text to image synthesis
 
Explainable Machine Learning (Explainable ML)
Explainable Machine Learning (Explainable ML)Explainable Machine Learning (Explainable ML)
Explainable Machine Learning (Explainable ML)
 
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAIGenerative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
 
Transfer learning-presentation
Transfer learning-presentationTransfer learning-presentation
Transfer learning-presentation
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial Networks
 
GAN - Theory and Applications
GAN - Theory and ApplicationsGAN - Theory and Applications
GAN - Theory and Applications
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networks
 
Machine learning in Cyber Security
Machine learning in Cyber SecurityMachine learning in Cyber Security
Machine learning in Cyber Security
 
Generative adversarial networks
Generative adversarial networksGenerative adversarial networks
Generative adversarial networks
 
Unsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGANUnsupervised learning represenation with DCGAN
Unsupervised learning represenation with DCGAN
 

Semelhante a Adversarial Attacks and Defenses in Deep Learning.pdf

Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...
Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...
Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...Kishor Datta Gupta
 
Adversarial Training is all you Need.pptx
Adversarial Training is all you Need.pptxAdversarial Training is all you Need.pptx
Adversarial Training is all you Need.pptxPrerana Khatiwada
 
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...Pluribus One
 
[Paper Review] Audio adversarial examples
[Paper Review] Audio adversarial examples[Paper Review] Audio adversarial examples
[Paper Review] Audio adversarial examplesJaeSung Bae
 
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...GeekPwn Keen
 
Bringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine LearningBringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine LearningBobby Filar
 
slides_security_and_privacy_in_machine_learning.pptx
slides_security_and_privacy_in_machine_learning.pptxslides_security_and_privacy_in_machine_learning.pptx
slides_security_and_privacy_in_machine_learning.pptxssuserabf73f
 
Survey of Adversarial Attacks in Deep Learning Models
Survey of Adversarial Attacks in Deep Learning ModelsSurvey of Adversarial Attacks in Deep Learning Models
Survey of Adversarial Attacks in Deep Learning ModelsIRJET Journal
 
Black-Box attacks against Neural Networks - technical project presentation
Black-Box attacks against Neural Networks - technical project presentationBlack-Box attacks against Neural Networks - technical project presentation
Black-Box attacks against Neural Networks - technical project presentationRoberto Falconi
 
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learning
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learningDEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learning
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learningFelipe Prado
 
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurity
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurityAI Cybersecurity: Pros & Cons. AI is reshaping cybersecurity
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurityTasnim Alasali
 
Machine Duping 101: Pwning Deep Learning Systems
Machine Duping 101: Pwning Deep Learning SystemsMachine Duping 101: Pwning Deep Learning Systems
Machine Duping 101: Pwning Deep Learning SystemsClarence Chio
 
Adversarial robustness using clever hans
Adversarial robustness using clever hansAdversarial robustness using clever hans
Adversarial robustness using clever hansShrutiMishra164
 
Attacks on Victim Model! A Defense Strategy
Attacks on Victim Model! A Defense StrategyAttacks on Victim Model! A Defense Strategy
Attacks on Victim Model! A Defense StrategySivaranjanikumar1
 
Robustness of compressed CNNs
Robustness of compressed CNNsRobustness of compressed CNNs
Robustness of compressed CNNsKaushalya Madhawa
 
Deceiving Autonomous Cars with Toxic Signs
Deceiving Autonomous Cars with Toxic SignsDeceiving Autonomous Cars with Toxic Signs
Deceiving Autonomous Cars with Toxic SignsLeonardoSalvucci1
 
Who is responsible for adversarial defense
Who is responsible for adversarial defenseWho is responsible for adversarial defense
Who is responsible for adversarial defenseKishor Datta Gupta
 
Securing Neural Networks
Securing Neural NetworksSecuring Neural Networks
Securing Neural NetworksTahseen Shabab
 
Subverting Machine Learning Detections for fun and profit
Subverting Machine Learning Detections for fun and profitSubverting Machine Learning Detections for fun and profit
Subverting Machine Learning Detections for fun and profitRam Shankar Siva Kumar
 

Semelhante a Adversarial Attacks and Defenses in Deep Learning.pdf (20)

Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...
Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...
Applicability issues of Evasion-Based Adversarial Attacks and Mitigation Tech...
 
Adversarial Training is all you Need.pptx
Adversarial Training is all you Need.pptxAdversarial Training is all you Need.pptx
Adversarial Training is all you Need.pptx
 
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...
Battista Biggio @ S+SSPR2014, Joensuu, Finland -- Poisoning Complete-Linkage ...
 
[Paper Review] Audio adversarial examples
[Paper Review] Audio adversarial examples[Paper Review] Audio adversarial examples
[Paper Review] Audio adversarial examples
 
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...
PRACTICAL ADVERSARIAL ATTACKS AGAINST CHALLENGING MODELS ENVIRONMENTS - Moust...
 
Bringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine LearningBringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine Learning
 
slides_security_and_privacy_in_machine_learning.pptx
slides_security_and_privacy_in_machine_learning.pptxslides_security_and_privacy_in_machine_learning.pptx
slides_security_and_privacy_in_machine_learning.pptx
 
Survey of Adversarial Attacks in Deep Learning Models
Survey of Adversarial Attacks in Deep Learning ModelsSurvey of Adversarial Attacks in Deep Learning Models
Survey of Adversarial Attacks in Deep Learning Models
 
Black-Box attacks against Neural Networks - technical project presentation
Black-Box attacks against Neural Networks - technical project presentationBlack-Box attacks against Neural Networks - technical project presentation
Black-Box attacks against Neural Networks - technical project presentation
 
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learning
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learningDEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learning
DEF CON 27 - workshop - YACIN NADJI - hands on adverserial machine learning
 
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurity
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurityAI Cybersecurity: Pros & Cons. AI is reshaping cybersecurity
AI Cybersecurity: Pros & Cons. AI is reshaping cybersecurity
 
Adversarial ML - Part 2.pdf
Adversarial ML - Part 2.pdfAdversarial ML - Part 2.pdf
Adversarial ML - Part 2.pdf
 
Machine Duping 101: Pwning Deep Learning Systems
Machine Duping 101: Pwning Deep Learning SystemsMachine Duping 101: Pwning Deep Learning Systems
Machine Duping 101: Pwning Deep Learning Systems
 
Adversarial robustness using clever hans
Adversarial robustness using clever hansAdversarial robustness using clever hans
Adversarial robustness using clever hans
 
Attacks on Victim Model! A Defense Strategy
Attacks on Victim Model! A Defense StrategyAttacks on Victim Model! A Defense Strategy
Attacks on Victim Model! A Defense Strategy
 
Robustness of compressed CNNs
Robustness of compressed CNNsRobustness of compressed CNNs
Robustness of compressed CNNs
 
Deceiving Autonomous Cars with Toxic Signs
Deceiving Autonomous Cars with Toxic SignsDeceiving Autonomous Cars with Toxic Signs
Deceiving Autonomous Cars with Toxic Signs
 
Who is responsible for adversarial defense
Who is responsible for adversarial defenseWho is responsible for adversarial defense
Who is responsible for adversarial defense
 
Securing Neural Networks
Securing Neural NetworksSecuring Neural Networks
Securing Neural Networks
 
Subverting Machine Learning Detections for fun and profit
Subverting Machine Learning Detections for fun and profitSubverting Machine Learning Detections for fun and profit
Subverting Machine Learning Detections for fun and profit
 

Último

Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfOrbitshub
 
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...apidays
 
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot ModelMcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot ModelDeepika Singh
 
Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..Christopher Logan Kennedy
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FMESafe Software
 
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ..."I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...Zilliz
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Orbitshub
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusZilliz
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDropbox
 
Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Zilliz
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAndrey Devyatkin
 
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamDEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamUiPathCommunity
 
FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024The Digital Insurer
 
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...apidays
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Victor Rentea
 
Artificial Intelligence Chap.5 : Uncertainty
Artificial Intelligence Chap.5 : UncertaintyArtificial Intelligence Chap.5 : Uncertainty
Artificial Intelligence Chap.5 : UncertaintyKhushali Kathiriya
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingEdi Saputra
 
WSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering DevelopersWSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering DevelopersWSO2
 
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Angeliki Cooney
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoffsammart93
 

Último (20)

Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdfRising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
Rising Above_ Dubai Floods and the Fortitude of Dubai International Airport.pdf
 
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
Apidays New York 2024 - Passkeys: Developing APIs to enable passwordless auth...
 
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot ModelMcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot Model
Mcleodganj Call Girls đŸ„° 8617370543 Service Offer VIP Hot Model
 
Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..Understanding the FAA Part 107 License ..
Understanding the FAA Part 107 License ..
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
 
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ..."I see eyes in my soup": How Delivery Hero implemented the safety system for ...
"I see eyes in my soup": How Delivery Hero implemented the safety system for ...
 
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
 
Exploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with MilvusExploring Multimodal Embeddings with Milvus
Exploring Multimodal Embeddings with Milvus
 
DBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor PresentationDBX First Quarter 2024 Investor Presentation
DBX First Quarter 2024 Investor Presentation
 
Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)Introduction to Multilingual Retrieval Augmented Generation (RAG)
Introduction to Multilingual Retrieval Augmented Generation (RAG)
 
AWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of TerraformAWS Community Day CPH - Three problems of Terraform
AWS Community Day CPH - Three problems of Terraform
 
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 AmsterdamDEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
DEV meet-up UiPath Document Understanding May 7 2024 Amsterdam
 
FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024FWD Group - Insurer Innovation Award 2024
FWD Group - Insurer Innovation Award 2024
 
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
Apidays New York 2024 - APIs in 2030: The Risk of Technological Sleepwalk by ...
 
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024Finding Java's Hidden Performance Traps @ DevoxxUK 2024
Finding Java's Hidden Performance Traps @ DevoxxUK 2024
 
Artificial Intelligence Chap.5 : Uncertainty
Artificial Intelligence Chap.5 : UncertaintyArtificial Intelligence Chap.5 : Uncertainty
Artificial Intelligence Chap.5 : Uncertainty
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
 
WSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering DevelopersWSO2's API Vision: Unifying Control, Empowering Developers
WSO2's API Vision: Unifying Control, Empowering Developers
 
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
Biography Of Angeliki Cooney | Senior Vice President Life Sciences | Albany, ...
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
 

Adversarial Attacks and Defenses in Deep Learning.pdf

  • 1. Adversarial Attacks and Defenses in Deep Learning Kui Ren, Tiahnhang Zheng, Zhan Qin, Xue Liu (2020) Engineering
  • 2. On April 23, 2013, Syrian hackers compromised the Associated Press Twitter feed and tweeted, “Breaking: Two Explosions in the White House and Barack Obama is injured”. In response to the tweet, the Dow Jones Industrial Average dropped by $136 billion dollars (although this drop was reversed 3 minutes later).
  • 3. What are adversarial attacks and why should you care? ● Any attempt to fool a deep learning model with deceptive input ● Especially researched in image recognition, but can also be applied to audio, text or tabular data ● When building models, we mostly focus on classiïŹcation effectiveness/ minimizing error. Relatively little work on model security and robustness. ● Imperceptible amounts of non-random noise can fool neural networks! ● Some of these attacks are 100% effective in fooling normal neural networks!
  • 4. Gong & Poellabauer, 2018. Comp Sci & Engineering
  • 5.
  • 6.
  • 7. What I’ll talk about ● Threat-models ● Some background terminology ● Notable adversarial models ● Notable adversarial defences ● Trends and remaining challenges ● Code
  • 8. Level of threat ● White-box: full knowledge of model architecture and parameters ● Gray-box: knowledge limited to features, model type ● Black-box: no/minimal knowledge of the model, can only use output All non-adversarially trained models are susceptible, even to black box models Adversarially trained models are still susceptible to white box models
  • 9. Background Adversarial loss: J(Ξ, x, y), Ξ = model weights An adversarial sample x’ : D(x,x’) < η (predeïŹned distance constraint, perturbation) ● Idea: ïŹnd the minimum difference or perturbation f(x’) ≠ y’
  • 10. Adversarial samples should be indistinguishable from benign samples Distance metrics: ● L₂ distance: What is the squared difference between adversarial and benign image? ● L∞ distance: Maximum element-wise difference between adversarial and benign image (for each pixel, take the absolute value difference between X and Z, and return the largest such distance found over all pixels)
  • 11. Notable adversarial models Limited-memory BFGS Grid search / line search to ïŹnd optimal hyperparameter Carlini and Wagner (C&W) attack Set of optimization-based attacks that generate L₀, L₂ and L∞ norm measured adversarial samples, with some restrictions (kappa) to make sure a valid image is produced 100% attack success on ‘normal’ neural networks trained on MNIST, CIFAR-10, ImageNet Compromised defensive models
  • 12. Notable adversarial models DeepFool “Iterative linearization of the classiïŹer to generate minimal perturbations that are sufïŹcient to change classiïŹcation labels” Computes perturbations more reliably Moosavi-Dezfooli et al., https://arxiv.org/pdf/1511.04599.pdf
  • 13.
  • 14. Notable adversarial models Universal adversarial attack ● Is there a universal perturbation that will work on most samples? ● L-BFGS- based ● Effective in attacking NN like CaffeNet, GoogleNet, VGG and Resnet ● Fooling rate > 53%
  • 15. Adversarial patch Only certain areas are perturbed
  • 16. Text & Audio Models 1% audio perturbation can change 50 words in text transcription! Attacks are robust to MP3 compression, but get lost when played over speakers https://nicholas.carlini.com/code/audio_adversarial_examples/ Strategies for text attacks generally include deleting, inserting, and modifying characters/words
  • 17. Adversarial defenses fall into 5 categories 1. Training on adversarial samples 2. Randomization 3. Adding noise 4. Removing noise 5. Mathematically provable defenses
  • 18. Defang: Randomize input or features ● Randomly padding and resizing input; image transformations with randomness
  • 19. ● Add random noise layer before each convolutional layer in training and test sets (RSE) ● Random feature pruning at each layer
  • 20. Detect: Denoise the input or features ● Conventional input rectiïŹcation ○ ‘Squeeze’ image → if output is very different from input, then likely adversarial ● GAN-based ○ Use GAN to learn benign data distribution ○ Generate a benign projection for the adversarial sample ● Autoencoder-based ○ Detector & reformer ○ Use an autocoder to compress input and learn manifold of benign samples ○ Detector compares each sample to learnt manifold ○ Reformer rectiïŹes adversarial samples
  • 21. Detect: Denoise the input or features ● High-level representation guided denoiser (HGD) ○ Trains a denoising u-net using a feature-level loss function to minimize feature differences between benign and adversarial samples ○ Won ïŹrst place in black-box defenses, 2017 ○ Even so, certain (white-box) attacks can reduce effectiveness to 0%
  • 22. Provable (certiïŹcated) defenses ● Defenses that have theoretical backing to have a certain accuracy against attacks ● Range of defenses include KNN and Bayesian-based defenses ● Consistency-based defenses: ○ Perturbations also affect the area around them ○ > 90 detection rate ● Very computationally intensive
  • 23. Trends in adversarial research ● Design stronger attacks to probe for weaknesses ● Real-world attack capabilities ● CertiïŹcated defenses - but currently not scalable “A problem is that an attack can only target one category of defenses, but defenses are required to 
 be effective against all possible attack methods” ● Analyzing model robustness - mostly done on KKN and linear classiïŹers
  • 24. Unresolved challenges ● Causality ● Does a general robust decision boundary exist that could be learnt by (certain) neural networks? ● Effectiveness vs efïŹciency ○ Adversarial training is effective, but requires a lot of data and compute ○ Randomization and denoising strategies very efïŹcient, but not as effective as claimed
  • 25. Discussion In what other ways are models not robust? Is model robustness/ security applicable to what you do / to our students? Model fairness has been a hot topic lately, but robustness/ security seems to lag behind - what do you think needs to change for adversarial training to be widely implemented? What are your thoughts on the paper in general?
  • 26. Try it yourself Benchmark machine learning systems' vulnerability to adversarial examples: https://github.com/cleverhans-lab/cleverhans Blog: cleverhans.io