Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Wild patterns - Ten years after the rise of Adversarial Machine Learning - NeCS 2019

1,047 views

Published on

Slides from the talk given by Battista Biggio at NeCS PhD Winter School, February 2019, Fai della Paganella, Trento

Published in: Engineering
  • Real Money Streams ~ Create multiple streams of wealth from your home! ♣♣♣ http://ishbv.com/ezpayjobs/pdf
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here
  • ♣♣ 10 Easy Ways to Improve Your Performance in Bed... ▲▲▲ http://ishbv.com/rockhardx/pdf
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here

Wild patterns - Ten years after the rise of Adversarial Machine Learning - NeCS 2019

  1. 1. Wild Patterns: Ten Years After the Rise of Adversarial Machine Learning Battista Biggio Pattern Recognition and Applications Lab University of Cagliari, Italy NeCS PhD Winter School, Fai della Paganella, Trento, Italy, Feb. 20, 2019 * Slides from this talk are inspired from the tutorial I prepared with Fabio Roli on such topic. https://www.pluribus-one.it/research/sec-ml/wild-patterns/
  2. 2. http://pralab.diee.unica.it @biggiobattista A Question to Start… What is the oldest survey article on machine learning that you have ever read? What is the publication year? 2
  3. 3. http://pralab.diee.unica.it @biggiobattista This Is Mine… Year 1966 3Credits: Dr Gavin Brown for showing me this article
  4. 4. http://pralab.diee.unica.it @biggiobattista Applications in the Old Good Days… What applications do you think that this paper dealt with? 4
  5. 5. http://pralab.diee.unica.it @biggiobattista Popular Applications in the Sixties 5 OCR for bank cheque sorting Aerial photo recognition Detection of particle tracks in bubble chambers
  6. 6. http://pralab.diee.unica.it @biggiobattista Key Feature of these Apps 6 Specialised applications for professional users…
  7. 7. 7 What about Today Applications?
  8. 8. http://pralab.diee.unica.it @biggiobattista Computer Vision for Self-Driving Cars 8 He et al., Mask R-CNN, ICCV ’17, https://arxiv.org/abs/1703.06870 Video from: https://www.youtube.com/watch?v=OOT3UIXZztE
  9. 9. http://pralab.diee.unica.it @biggiobattista Automatic Speech Recognition for Virtual Assistants Amazon Alexa Apple Siri Microsoft Cortana Google Assistant 9
  10. 10. http://pralab.diee.unica.it @biggiobattista Today Applications of Machine Learning 10
  11. 11. http://pralab.diee.unica.it @biggiobattista Key Features of Today Apps 11 Personal and consumer applications…
  12. 12. http://pralab.diee.unica.it @biggiobattista We Are Living in the Best of the Worlds… 12 AI is going to transform industry and business as electricity did about a century ago (Andrew Ng, Jan. 2017)
  13. 13. 13 All Right? All Good?
  14. 14. http://pralab.diee.unica.it @biggiobattista iPhone 5s and 6s with Fingerprint Reader… 14
  15. 15. http://pralab.diee.unica.it @biggiobattista Hacked a Few Days After Release… 15
  16. 16. 16 But maybe this happens only for old, shallow machine learning… End-to-end deep learning is another story…
  17. 17. http://pralab.diee.unica.it @biggiobattista Adversarial School Bus 17 Szegedy et al., Intriguing properties of neural networks, ICLR 2014 Biggio, Roli et al., Evasion attacks against machine learning at test time, ECML-PKDD 2013
  18. 18. http://pralab.diee.unica.it @biggiobattista Adversarial Glasses • M. Sharif et al. (ACM CCS 2016) attacked deep neural networks for face recognition with carefully-fabricated eyeglass frames • When worn by a 41-year-old white male (left image), the glasses mislead the deep network into believing that the face belongs to the famous actress Milla Jovovich 18
  19. 19. 19 But maybe this happens only for image recognition…
  20. 20. http://pralab.diee.unica.it @biggiobattista Audio Adversarial Examples 20 “without the dataset the article is useless” “okay google browse to evil dot com” Transcription by Mozilla DeepSpeechAudio https://nicholas.carlini.com/code/audio_adversarial_examples/
  21. 21. http://pralab.diee.unica.it @biggiobattista Deep Neural Networks for EXE Malware Detection • MalConv: convolutional deep network trained on raw bytes to detect EXE malware • Gradient-based attacks can evade it by adding few padding bytes 21[Kolosniaji, Biggio, Roli et al., Adversarial Malware Binaries, EUSIPCO2018]
  22. 22. http://pralab.diee.unica.it @biggiobattista Take-home Message We are living exciting time for machine learning… …Our work feeds a lot of consumer technologies for personal applications... This opens up new big possibilities, but also new security risks 22
  23. 23. 23 Where Do These Security Risks Come From?
  24. 24. http://pralab.diee.unica.it @biggiobattista The Classical Statistical Model Note these two implicit assumptions of the model: 1. The source of data is given, and it does not depend on the classifier 2. Noise affecting data is stochastic 24 Data source acquisition/ measurementRaw data x1 x2 ... xd feature vector learning algorithm classifier stochastic noise ed by sets of coupled s for formal neurons ation of essentials feat Example: OCR
  25. 25. http://pralab.diee.unica.it @biggiobattista Can This Model Be Used Under Attack? 25 Data source acquisition/ measurementRaw data x1 x2 ... xd feature vector learning algorithm classifier stochastic noise ed by sets of coupled s for formal neurons ation of essentials feat Example: OCR
  26. 26. http://pralab.diee.unica.it @biggiobattista An Example: Spam Filtering 26 Total score = 6.0 From: spam@example.it Buy Viagra ! > 5.0 (threshold) Spam Linear Classifier ØThe famous SpamAssassin filter is really a linear classifier §http://spamassassin.apache.org Feature weights buy = 1.0 viagra = 5.0
  27. 27. http://pralab.diee.unica.it @biggiobattista Feature Space View 27 X2 X1 + + + + + - - - - - yc(x) Feature weights buy = 1.0 viagra = 5.0 From: spam@example.it Buy Viagra! • Classifier’s weights are learned from training data • The SpamAssassin filter uses the perceptron algorithm
  28. 28. 28 But spam filtering is not a stationary classification task, the data source is not neutral…
  29. 29. http://pralab.diee.unica.it @biggiobattista The Data Source Can Add “Good” Words 29 Total score = 1.0 From: spam@example.it Buy Viagra ! conference meeting < 5.0 (threshold) Linear Classifier Feature weights buy = 1.0 viagra = 5.0 conference = -2.0 meeting = -3.0 Ham üAdding good words is a typical spammers trick [Z. Jorgensen et al., JMLR 2008]
  30. 30. http://pralab.diee.unica.it @biggiobattista Adding Good Words: Feature Space View 30 X2 X1 + + + + + - - - - - yc(x) Feature weights buy = 1.0 viagra = 5.0 conference = -2.0 meeting = -3.0 From: spam@example.it Buy Viagra! conference meeting - üNote that spammers corrupt patterns with a noise that is not random..
  31. 31. http://pralab.diee.unica.it @biggiobattista Is This Model Good for Spam Filtering? Ø The source of data is given, and it does not depend on the classifier Ø Noise affecting data is stochastic (“random”) 31 Data source acquisition/ measurementRaw data x1 x2 ... xd feature vector learning algorithm classifier stochastic noise ed by sets of coupled s for formal neurons ation of essentials feat Example: OCR
  32. 32. 32 No, it is not…
  33. 33. http://pralab.diee.unica.it @biggiobattista Adversarial Machine Learning 1. The source of data is not neutral, it depends on the classifier 2. Noise is not stochastic, it is adversarial, crafted to maximize the classification error 33 measurementRaw data x1 x2 ... xn feature vector learning algorithm classifier adversarial noise Spam message: Buy Viagra Camouflaged message: Buy Vi@gra Dublin University Non-neutral data source
  34. 34. http://pralab.diee.unica.it @biggiobattista Adversarial Noise vs. Stochastic Noise 34 Hamming’s adversarial noise model: the channel acts as an adversary that arbitrarily corrupts the code-word subject to a bound on the total number of errors Shannon’s stochastic noise model: probabilistic model of the channel, the probability of occurrence of too many or too few errors is usually low • This distinction is not new...
  35. 35. http://pralab.diee.unica.it @biggiobattista The Classical Model Cannot Work • Standard classification algorithms assume that – data generating process is independent from the classifier – training /test data follow the same distribution (i.i.d. samples) • This is not the case for adversarial tasks! • Easy to see that classifier performance will degrade quickly if the adversarial noise is not taken into account – Adversarial tasks are a mission impossible for the classical model 35
  36. 36. 36 How Should We Design Pattern Classifiers Under Attack?
  37. 37. http://pralab.diee.unica.it @biggiobattista Adversary-aware Machine Learning 37 Machine learning systems should be aware of the arms race with the adversary [Biggio, Fumera, Roli. Security evaluation of pattern classifiers under attack, IEEE TKDE, 2014] Adversary System Designer Analyze system Devise and execute attack Analyze attack Develop countermeasure
  38. 38. http://pralab.diee.unica.it @biggiobattista • In 2004 spammers invented a new trick for evading anti-spam filters… – As filters did not analyze the content of attached images… – Spammers embedded their messages into images…so evading filters… Image-based Spam Your orological prescription appointment starts September 30th bergstrom mustsquawbush try bimini , maine see woodwind in con or patagonia or scrapbook but. patriarchal and tasteful must advisory not thoroughgoing the frowzy not ellwood da jargon and. beresford ! arpeggio must stern try disastrous ! alone , wear da esophagi try autonomic da clyde and taskmaster, tideland try cream see await must mort in. From: Conrad Stern <rjlfm@berlin.de> To: utente@emailserver.it Arms Race: The Case of Image Spam 38
  39. 39. http://pralab.diee.unica.it @biggiobattista • PRA Lab team proposed a countermeasure against image spam… – G. Fumera, I. Pillai, F. Roli, Spam filtering based on the analysis of text information embedded into images, Journal of Machine Learning Research, Vol. 7, 2006 • Text embedded in images is read by Optical Character Recognition (OCR) • OCRing image text and fusing it with other mail data allows discriminating spam/ham mails Arms Race: The Case of Image Spam 39
  40. 40. http://pralab.diee.unica.it @biggiobattista • The OCR-based solution was deployed as a plug-in of SpamAssassin filter (called Bayes OCR) and worked well for a while… http://wiki.apache.org/spamassassin/CustomPlugins Arms Race: The Case of Image Spam 40
  41. 41. http://pralab.diee.unica.it @biggiobattista Spammers’ Reaction • Spammers reacted quickly with a countermeasure against OCR-based solutions (and against signature-based image spam detection) • They applied content obscuring techniques to images, like done in CAPTCHAs, to make OCR systems ineffective without compromising human readability 41
  42. 42. http://pralab.diee.unica.it @biggiobattista You find the complete story here: http://en.wikipedia.org/wiki/Image_spam Arms Race: The Case of Image Spam • PRA Lab did another countermove by devising features which detect the presence of spammers obfuscation techniques in text images ü A feature for detecting characters fragmented or mixed with small background components ü A feature for detecting characters connected through background components ü A feature for detecting non-uniform background, hidden text • This solution was deployed as a new plug-in of SpamAssassin filter (called Image Cerberus) 42
  43. 43. 43 How Can We Design Adversary-aware Machine Learning Systems?
  44. 44. http://pralab.diee.unica.it @biggiobattista Adversary-aware Machine Learning 44 Machine learning systems should be aware of the arms race with the adversary [Biggio, Fumera, Roli. Security evaluation of pattern classifiers under attack, IEEE TKDE, 2014] Adversary System Designer Analyze system Devise and execute attack Analyze attack Develop countermeasure
  45. 45. http://pralab.diee.unica.it @biggiobattista Adversary-aware Machine Learning 45 Machine learning systems should be aware of the arms race with the adversary [Biggio, Fumera, Roli. Security evaluation of pattern classifiers under attack, IEEE TKDE, 2014] System Designer System Designer Simulate attack Evaluate attack’s impact Develop countermeasureModel adversary
  46. 46. http://pralab.diee.unica.it @biggiobattista The Three Golden Rules 1. Know your adversary 2. Be proactive 3. Protect your classifier 46
  47. 47. 47 Know your adversary If you know the enemy and know yourself, you need not fear the result of a hundred battles (Sun Tzu, The art of war, 500 BC)
  48. 48. http://pralab.diee.unica.it @biggiobattista Adversary’s 3D Model 48 Adversary’s Knowledge Adversary’s Capability Adversary’s Goal
  49. 49. http://pralab.diee.unica.it @biggiobattista Adversary’s Goal • To cause a security violation... 49 Misclassifications that do not compromise normal system operation Integrity Misclassifications that compromise normal system operation (denial of service) Availability Querying strategies that reveal confidential information on the learning model or its users Confidentiality / Privacy [Barreno et al., Can Machine Learning Be Secure? ASIACCS ‘06]
  50. 50. http://pralab.diee.unica.it @biggiobattista Adversary’s Knowledge • Perfect-knowledge (white-box) attacks – upper bound on the performance degradation under attack 50 TRAINING DATA FEATURE REPRESENTATION LEARNING ALGORITHM e.g., SVM x1 x2 ... xd x xx x x x x x x x x x x xxx x - Learning algorithm - Parameters (e.g., feature weights) - Feedback on decisions [B. Biggio, G. Fumera, F. Roli, IEEE TKDE 2014]
  51. 51. http://pralab.diee.unica.it @biggiobattista • Limited-knowledge Attacks – Ranging from gray-box to black-box attacks 51 TRAINING DATA FEATURE REPRESENTATION LEARNING ALGORITHM e.g., SVM x1 x2 ... xd x xx x x x x x x x x x x xxx x - Learning algorithm - Parameters (e.g., feature weights) - Feedback on decisions Adversary’s Knowledge [B. Biggio, G. Fumera, F. Roli, IEEE TKDE 2014]
  52. 52. http://pralab.diee.unica.it @biggiobattista Kerckhoffs’ Principle • Kerckhoffs’ Principle (Kerckhoffs 1883) states that the security of a system should not rely on unrealistic expectations of secrecy – It’s the opposite of the principle of “security by obscurity” • Secure systems should make minimal assumptions about what can realistically be kept secret from a potential attacker • For machine learning systems, one could assume that the adversary is aware of the learning algorithm and can obtain some degree of information about the data used to train the learner • But the best strategy is to assess system security under different levels of adversary’s knowledge 52[Joseph et al., Adversarial Machine Learning, Cambridge Univ. Press, 2017]
  53. 53. http://pralab.diee.unica.it @biggiobattista Adversary’s Capability 53[B. Biggio, G. Fumera, F. Roli, IEEE TKDE 2014; M. Barreno et al., ML 2010] • Attackers may manipulate training data and/or test data TRAINING TEST Influence model at training time to cause subsequent errors at test time poisoning attacks, backdoors Manipulate malicious samples at test time to cause misclassications evasion attacks, adversarial examples
  54. 54. http://pralab.diee.unica.it @biggiobattista A Deliberate Poisoning Attack? 54 [http://exploringpossibilityspace.blogspot.it/2016 /03/poor-software-qa-is-root-cause-of-tay.html] Microsoft deployed Tay, and AI chatbot designed to talk to youngsters on Twitter, but after 16 hours the chatbot was shut down since it started to raise racist and offensive comments.
  55. 55. http://pralab.diee.unica.it @biggiobattista Adversary’s Capability • Luckily, the adversary is not omnipotent, she is constrained… 55[R. Lippmann, Dagstuhl Workshop, 2012] Email messages must be understandable by human readers Malware must execute on a computer, usually exploiting a known vulnerability
  56. 56. http://pralab.diee.unica.it @biggiobattista • Constraints on data manipulation – maximum number of samples that can be added to the training data • the attacker usually controls only a small fraction of the training samples – maximum amount of modifications • application-specific constraints in feature space • e.g., max. number of words that are modified in spam emails 56 d(x, !x ) ≤ dmax x2 x1 f(x) x Feasible domain x ' TRAINING TEST Adversary’s Capability [B. Biggio, G. Fumera, F. Roli, IEEE TKDE 2014]
  57. 57. http://pralab.diee.unica.it @biggiobattista Conservative Design • The design and analysis of a system should avoid unnecessary or unreasonable assumptions on the adversary’s capability – worst-case security evaluation • Conversely, analysing the capabilities of an omnipotent adversary reveals little about a learning system’s behaviour against realistically-constrained attackers • Again, the best strategy is to assess system security under different levels of adversary’s capability 57[Joseph et al., Adversarial Machine Learning, Cambridge Univ. Press, 2017]
  58. 58. 58 Be Proactive To know your enemy, you must become your enemy (Sun Tzu, The art of war, 500 BC)
  59. 59. http://pralab.diee.unica.it @biggiobattista Be Proactive • Given a model of the adversary characterized by her: – Goal – Knowledge – Capability Try to anticipate the adversary! • What is the optimal attack she can do? • What is the expected performance decrease of your classifier? 59
  60. 60. http://pralab.diee.unica.it @biggiobattista Evasion of Linear Classifiers • Problem: how to evade a linear (trained) classifier? Start 2007 with a bang! Make WBFS YOUR PORTFOLIO’s first winner of the year ... start bang portfolio winner year ... university campus 1 1 1 1 1 ... 0 0 +6 > 0, SPAM (correctly classified) f (x) = sign(wT x) x start bang portfolio winner year ... university campus +2 +1 +1 +1 +1 ... -3 -4 w x’ St4rt 2007 with a b4ng! Make WBFS YOUR PORTFOLIO’s first winner of the year ... campus start bang portfolio winner year ... university campus 0 0 1 1 1 ... 0 1 +3 -4 < 0, HAM (misclassified email) f (x) = sign(wT x) 60
  61. 61. http://pralab.diee.unica.it @biggiobattista Evasion of Nonlinear Classifiers • What if the classifier is nonlinear? • Decision functions can be arbitrarily complicated, with no clear relationship between features (x) and classifier parameters (w) −2−1.5−1−0.500.511.5 61
  62. 62. http://pralab.diee.unica.it @biggiobattista Detection of Malicious PDF Files Srndic & Laskov, Detection of malicious PDF files based on hierarchical document structure, NDSS 2013 “The most aggressive evasion strategy we could conceive was successful for only 0.025% of malicious examples tested against a nonlinear SVM classifier with the RBF kernel [...]. Currently, we do not have a rigorous mathematical explanation for such a surprising robustness. Our intuition suggests that [...] the space of true features is “hidden behind” a complex nonlinear transformation which is mathematically hard to invert. [...] the same attack staged against the linear classifier [...] had a 50% success rate; hence, the robustness of the RBF classifier must be rooted in its nonlinear transformation” 62
  63. 63. http://pralab.diee.unica.it @biggiobattista Evasion Attacks against Machine Learning at Test Time Biggio, Corona, Maiorca, Nelson, Srndic, Laskov, Giacinto, Roli, ECML-PKDD 2013 • Goal: maximum-confidence evasion • Knowledge: perfect (white-box attack) • Attack strategy: • Non-linear, constrained optimization – Projected gradient descent: approximate solution for smooth functions • Gradients of g(x) can be analytically computed in many cases – SVMs, Neural networks −2−1.5−1−0.500.51 x f (x) = sign g(x)( )= +1, malicious −1, legitimate " # $ %$ x ' min $% &((% ) s. t. ( − (% . ≤ 0123 63
  64. 64. http://pralab.diee.unica.it @biggiobattista Computing Descent Directions Support vector machines Neural networks x1 xd d1 dk dm xf g(x) w1 wk wm v11 vmd vk1 …… …… g(x) = αi yik(x, i ∑ xi )+ b, ∇g(x) = αi yi∇k(x, xi ) i ∑ g(x) = 1+exp − wkδk (x) k=1 m ∑ # $ % & ' ( ) * + , - . −1 ∂g(x) ∂xf = g(x) 1− g(x)( ) wkδk (x) 1−δk (x)( )vkf k=1 m ∑ RBF kernel gradient: ∇k(x,xi ) = −2γ exp −γ || x − xi ||2 { }(x − xi ) [Biggio, Roli et al., ECML PKDD 2013] 64
  65. 65. http://pralab.diee.unica.it @biggiobattista An Example on Handwritten Digits • Nonlinear SVM (RBF kernel) to discriminate between ‘3’ and ‘7’ • Features: gray-level pixel values (28 x 28 image = 784 features) Few modifications are enough to evade detection! 1st adversarial examples generated with gradient-based attacks date back to 2013! (one year before attacks to deep neural networks) Before attack (3 vs 7) 5 10 15 20 25 5 10 15 20 25 After attack, g(x)=0 5 10 15 20 25 5 10 15 20 25 After attack, last iter. 5 10 15 20 25 5 10 15 20 25 0 500 −2 −1 0 1 2 g(x) number of iterations After attack (misclassified as 7) 65[Biggio, Roli et al., ECML PKDD 2013]
  66. 66. http://pralab.diee.unica.it @biggiobattista Bounding the Adversary’s Knowledge Limited-knowledge (gray/black-box) attacks • Only feature representation and (possibly) learning algorithm are known • Surrogate data sampled from the same distribution as the classifier’s training data • Classifier’s feedback to label surrogate data PD(X,Y)data Surrogate training data Send queries Get labels f(x) Learn surrogate classifier f’(x) This is the same underlying idea behind substitute models and black- box attacks (transferability) [N. Papernot et al., IEEE Euro S&P ’16; N. Papernot et al., ASIACCS’17] 66[Biggio, Roli et al., ECML PKDD 2013]
  67. 67. http://pralab.diee.unica.it @biggiobattista Recent Results on Android Malware Detection • Drebin: Arp et al., NDSS 2014 – Android malware detection directly on the mobile phone – Linear SVM trained on features extracted from static code analysis [Demontis, Biggio et al., IEEE TDSC 2017] x2 Classifier 0 1 ... 1 0 Android app (apk) malware benign x1 x f (x) 67
  68. 68. http://pralab.diee.unica.it @biggiobattista Recent Results on Android Malware Detection • Dataset (Drebin): 5,600 malware and 121,000 benign apps (TR: 30K, TS: 60K) • Detection rate at FP=1% vs max. number of manipulated features (averaged on 10 runs) – Perfect knowledge (PK) white-box attack; Limited knowledge (LK) black-box attack 68[Demontis, Biggio et al., IEEE TDSC 2017]
  69. 69. http://pralab.diee.unica.it @biggiobattista Take-home Messages • Linear and non-linear supervised classifiers can be highly vulnerable to well-crafted evasion attacks • Performance evaluation should be always performed as a function of the adversary’s knowledge and capability – Security Evaluation Curves 69 min x' g(x') s.t. d(x, x') ≤ dmax x ≤ x'
  70. 70. http://pralab.diee.unica.it @biggiobattista Why Is Machine Learning So Vulnerable? • Learning algorithms tend to overemphasize some features to discriminate among classes • Large sensitivity to changes of such input features: !"#(") • Different classifiers tend to find the same set of relevant features – that is why attacks can transfer across models! 70[Melis et al., Explaining Black-box Android Malware Detection, EUSIPCO 2018] & g(&) &’ positivenegative
  71. 71. 71 2014: Deep Learning Meets Adversarial Machine Learning
  72. 72. http://pralab.diee.unica.it @biggiobattista The Discovery of Adversarial Examples 72 ... we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extent. We can cause the network to misclassify an image by applying a certain hardly perceptible perturbation, which is found by maximizing the network’s prediction error ... [Szegedy, Goodfellow et al., Intriguing Properties of NNs, ICLR 2014, ArXiv 2013]
  73. 73. http://pralab.diee.unica.it @biggiobattista Adversarial Examples and Deep Learning • C. Szegedy et al. (ICLR 2014) independently developed a gradient-based attack against deep neural networks – minimally-perturbed adversarial examples 73[Szegedy, Goodfellow et al., Intriguing Properties of NNs, ICLR 2014, ArXiv 2013]
  74. 74. http://pralab.diee.unica.it @biggiobattista !(#) ≠ & School Bus (x) Ostrich Struthio Camelus Adversarial Noise (r) The adversarial image x + r is visually hard to distinguish from x Informally speaking, the solution x + r is the closest image to x classified as l by f The solution is approximated using using a box-constrained limited-memory BFGS Creation of Adversarial Examples 74[Szegedy, Goodfellow et al., Intriguing Properties of NNs, ICLR 2014, ArXiv 2013]
  75. 75. Many Black Swans After 2014… [Search https://arxiv.org with keywords “adversarial examples”] 75 • Several defenses have been proposed against adversarial examples, and more powerful attacks have been developed to show that they are ineffective. Remember the arms race? • Most of these attacks are modifications to the optimization problems reported for evasion attacks / adversarial examples, using different gradient-based solution algorithms, initializations and stopping conditions. • Most popular attack algorithms: FGSM (Goodfellow et al.), JSMA (Papernot et al.), CW (Carlini & Wagner, and follow- up versions)
  76. 76. 76 Why Adversarial Perturbations are Imperceptible?
  77. 77. http://pralab.diee.unica.it @biggiobattista Why Adversarial Perturbations against Deep Networks are Imperceptible? • Large sensitivity of g(x) to input changes – i.e., the input gradient !"#(") has a large norm (scales with input dimensions!) – Thus, even small modifications along that direction will cause large changes in the predictions 77 [Simon-Gabriel et al., Adversarial vulnerability of NNs increases with input dimension, arXiv 2018] & g(&) &’ !"#(")
  78. 78. http://pralab.diee.unica.it @biggiobattista • Regularization also impacts (reduces) the size of input gradients – High regularization requires larger perturbations to mislead detection – e.g., see manipulated digits 9 (classified as 8) against linear SVMs with different C values Adversarial Perturbations and Regularization 78[Demontis, Biggio, Roli et al., Why Do Adversarial Attacks Transfer? ... arXiv 2018] high regularization large perturbation low regularization imperceptible perturbation C=0.001, eps=1.7 C=1.0, eps=0.47
  79. 79. 79 Is Deep Learning Safe for Robot Vision?
  80. 80. http://pralab.diee.unica.it @biggiobattista Is Deep Learning Safe for Robot Vision? • Evasion attacks against the iCub humanoid robot – Deep Neural Network used for visual object recognition 80[Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  81. 81. http://pralab.diee.unica.it @biggiobattista [http://old.iit.it/projects/data-sets] iCubWorld28 Data Set: Example Images 81
  82. 82. http://pralab.diee.unica.it @biggiobattista From Binary to Multiclass Evasion • In multiclass problems, classification errors occur in different classes. • Thus, the attacker may aim: 1. to have a sample misclassified as any class different from the true class (error-generic attacks) 2. to have a sample misclassified as a specific class (error-specific attacks) 82 cup sponge dish detergent Error-generic attacks any class cup sponge dish detergent Error-specific attacks [Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  83. 83. http://pralab.diee.unica.it @biggiobattista Error-generic Evasion • Error-generic evasion – k is the true class (blue) – l is the competing (closest) class in feature space (red) • The attack minimizes the objective to have the sample misclassified as the closest class (could be any!) 83 1 0 1 1 0 1 Indiscriminate evasion [Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  84. 84. http://pralab.diee.unica.it @biggiobattista • Error-specific evasion – k is the target class (green) – l is the competing class (initially, the blue class) • The attack maximizes the objective to have the sample misclassified as the target class Error-specific Evasion 84 max 1 0 1 1 0 1 Targeted evasion [Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  85. 85. http://pralab.diee.unica.it @biggiobattista Adversarial Examples against iCub – Gradient Computation ∇fi (x) = ∂fi(z) ∂z ∂z ∂x f1 f2 fi fc ... ... 85 The given optimization problems can be both solved with gradient-based algorithms The gradient of the objective can be computed using the chain rule 1. the gradient of the functions fi(z) can be computed if the chosen classifier is differentiable 2. ... and then backpropagated through the deep network with automatic differentiation
  86. 86. http://pralab.diee.unica.it @biggiobattista Example of Adversarial Images against iCub An adversarial example from class laundry-detergent, modified by the proposed algorithm to be misclassified as cup 86[Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  87. 87. http://pralab.diee.unica.it @biggiobattista The “Sticker” Attack against iCub Adversarial example generated by manipulating only a specific region, to simulate a sticker that could be applied to the real-world object. This image is classified as cup. 87[Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  88. 88. 88 Countering Evasion Attacks What is the rule? The rule is protect yourself at all times (from the movie “Million dollar baby”, 2004)
  89. 89. http://pralab.diee.unica.it @biggiobattista Security Measures against Evasion Attacks 1. Reduce sensitivity to input changes with robust optimization – Adversarial Training / Regularization 2. Introduce rejection / detection of adversarial examples 89 min $ ∑& max ||*+||,- ℓ(0&, 2$ 3& + *& ) bounded perturbation! 1 0 1 1 0 1 SVM-RBF (higher rejection rate) 1 0 1 1 0 1 SVM-RBF (no reject)
  90. 90. 90 Countering Evasion: Reducing Sensitivity to Input Changes with Robust Optimization
  91. 91. http://pralab.diee.unica.it @biggiobattista • Robust optimization (a.k.a. adversarial training) • Robustness and regularization (Xu et al., JMLR 2009) – under linearity of ℓ and "#, equivalent to robust optimization Reducing Input Sensitivity via Robust Optimization min # max ||*+||,-. ∑0 ℓ 10, "# 30 + *0 bounded perturbation! 91 min # ∑5 ℓ 10, "# 30 + 6||73"||8 dual norm of the perturbation ||73"||8 = ||#||8
  92. 92. http://pralab.diee.unica.it @biggiobattista Experiments on Android Malware • Infinity-norm regularization is the optimal regularizer against sparse evasion attacks – Sparse evasion attacks penalize | " |# promoting the manipulation of only few features Results on Adversarial Android Malware [Demontis, Biggio et al., Yes, ML Can Be More Secure!..., IEEE TDSC 2017] Absolute weight values |$| in descending order Why? It bounds the maximum weight absolute values! min w,b w ∞ +C max 0,1− yi f (xi )( ) i ∑ , w ∞ = max i=1,...,d wiSec-SVM 92
  93. 93. http://pralab.diee.unica.it @biggiobattista Adversarial Training and Regularization • Adversarial training can also be seen as a form of regularization, which penalizes the (dual) norm of the input gradients ! |#$ℓ |& • Known as double backprop or gradient/Jacobian regularization – see, e.g., Simon-Gabriel et al., Adversarial vulnerability of neural networks increases with input dimension, ArXiv 2018; and Lyu et al., A unified gradient regularization family for adversarial examples, ICDM 2015. 93 ' g(') '’ with adversarial trainingTake-home message: the net effect of these techniques is to make the prediction function of the classifier smoother (increasing the input margin)
  94. 94. http://pralab.diee.unica.it @biggiobattista Ineffective Defenses: Obfuscated Gradients • Work by Carlini & Wagner (SP’ 17) and Athalye et al. (ICML ‘18) has shown that – some recently-proposed defenses rely on obfuscated / masked gradients, and – they can be circumvented 94 g(") "’" Obfuscated gradients do not allow the correct execution of gradient-based attacks... " g(") "’ ... but substitute models and/or smoothing can correctly reveal meaningful input gradients!
  95. 95. 95 Countering Evasion: Detecting & Rejecting Adversarial Examples
  96. 96. http://pralab.diee.unica.it @biggiobattista Detecting & Rejecting Adversarial Examples • Adversarial examples tend to occur in blind spots – Regions far from training data that are anyway assigned to ‘legitimate’ classes 96 blind-spot evasion (not even required to mimic the target class) rejection of adversarial examples through enclosing of legitimate classes
  97. 97. http://pralab.diee.unica.it @biggiobattista Detecting & Rejecting Adversarial Examples input perturbation (Euclidean distance) 97[Melis, Biggio et al., Is Deep Learning Safe for Robot Vision? ICCVW ViPAR 2017]
  98. 98. http://pralab.diee.unica.it @biggiobattista [S. Sabour at al., ICLR 2016] Why Rejection (in Representation Space) Is Not Enough? 98
  99. 99. 99 Adversarial Examples and Security Evaluation (Demo Session)
  100. 100. http://pralab.diee.unica.it @biggiobattista secml: An open source Python library for ML Security 100 adv ml expl others - ML algorithms via sklearn - DL algorithms and optimizers via PyTorch and Tensorflow - attacks (evasion, poisoning, ...) with custom/faster solvers - defenses (advx rejection, adversarial training, ...) - Explanation methods based on influential features - Explanation methods based on influential prototypes - Parallel computation - Support for dense/sparse data - Advanced plotting functions (via matplotlib) - Modular and easy to extend First release scheduled on April 2019! Marco Melis Ambra Demontis
  101. 101. http://pralab.diee.unica.it @biggiobattista ML Security Evaluation Security Evaluation (Attack Simulation) attack model and parameters Security Risk Level (numeric value) machine-learning model secml data 101 e.g., evasion attacks – modify samples of class +1 with l1 perturbation, for eps in {0, 5, 10, 50} e.g., poisoning attacks – inject samples of classes +1/-1 in the training set, with % of poisoning points in {0, 2, 5, 10, 20}
  102. 102. http://pralab.diee.unica.it @biggiobattista Security Evaluation Curves • Security evaluation curves – accuracy vs increasing perturbation • Security value: – mean accuracy under attack • Security level: – Low / Med / High 102
  103. 103. http://pralab.diee.unica.it @biggiobattista Security Evaluation Curves 103
  104. 104. 104 Interactive Demos Adversarial Examples: https://sec-ml.pluribus-one.it/demo/ Security Evaluation: https://advx-secml.pluribus-one.it
  105. 105. http://pralab.diee.unica.it @biggiobattista EU H2020 Project ALOHA • ALOHA – software framework for runtime-Adaptive and secure deep Learning On Heterogeneous Architectures • Project goal: to facilitate implementation of deep learning algorithms on heterogeneous low- energy computing platforms • Project website: www.aloha-h2020.eu • Pluribus One is in charge of evaluating and improving security of deep learning algorithms under attack 105 This project has received funding from the European Union’s Horizon 2020 Research and Innovation programme under Grant Agreement No. 780788
  106. 106. 106 Poisoning Machine Learning
  107. 107. http://pralab.diee.unica.it @biggiobattista Poisoning Machine Learning 107 x xx x x x x x x x x x x xxx x x1 x2 ... xd pre-processing and feature extraction training data (with labels) classifier learning start bang portfolio winner year ... university campus Start 2007 with a bang! Make WBFS YOUR PORTFOLIO’s first winner of the year ... start bang portfolio winner year ... university campus 1 1 1 1 1 ... 0 0 xSPAM start bang portfolio winner year ... university campus +2 +1 +1 +1 +1 ... -3 -4 w x x x x xx x x x x x x x x xx x classifier generalizes well on test data
  108. 108. http://pralab.diee.unica.it @biggiobattista Poisoning Machine Learning 108 x xx x x x x x x x x x x xxx x x1 x2 ... xd pre-processing and feature extraction corrupted training data classifier learning is compromised... Start 2007 with a bang! Make WBFS YOUR PORTFOLIO’s first winner of the year ... university campus... start bang portfolio winner year ... university campus 1 1 1 1 1 ... 1 1 xSPAM start bang portfolio winner year ... university campus +2 +1 +1 +1 +1 ... +1 +1 w x x x x xx x x x x x x x x xx x ... to maximize error on test data xx x poisoning data
  109. 109. http://pralab.diee.unica.it @biggiobattista • Goal: to maximize classification error • Knowledge: perfect / white-box attack • Capability: injecting poisoning samples into TR • Strategy: find an optimal attack point xc in TR that maximizes classification error xc classification error = 0.039classification error = 0.022 Poisoning Attacks against Machine Learning xc classification error as a function of xc [Biggio, Nelson, Laskov. Poisoning attacks against SVMs. ICML, 2012] 109
  110. 110. http://pralab.diee.unica.it @biggiobattista Poisoning is a Bilevel Optimization Problem • Attacker’s objective – to maximize generalization error on untainted data, w.r.t. poisoning point xc • Poisoning problem against (linear) SVMs: Loss estimated on validation data (no attack points!) Algorithm is trained on surrogate data (including the attack point) [Biggio, Nelson, Laskov. Poisoning attacks against SVMs. ICML, 2012] [Xiao, Biggio, Roli et al., Is feature selection secure against training data poisoning? ICML, 2015] [Munoz-Gonzalez, Biggio, Roli et al., Towards poisoning of deep learning..., AISec 2017] max $% & '()*, ,∗ s. t. ,∗ = argmin6 ℒ '89 ∪ ;<, =< , , max $% > ?@A B max(0,1 − =?,∗ ;? ) s. t. ,∗ = argminH,I A J HK H + C ∑O@A P max(0,1 − =O, ;O ) + C max(0,1 − =<, ;< ) 110
  111. 111. http://pralab.diee.unica.it @biggiobattista xc (0) xc Gradient-based Poisoning Attacks • Gradient is not easy to compute – The training point affects the classification function • Trick: – Replace the inner learning problem with its equilibrium (KKT) conditions – This enables computing gradient in closed form • Example for (kernelized) SVM – similar derivation for Ridge, LASSO, Logistic Regression, etc. 111 xc (0) xc [Biggio, Nelson, Laskov. Poisoning attacks against SVMs. ICML, 2012] [Xiao, Biggio, Roli et al., Is feature selection secure against training data poisoning? ICML, 2015]
  112. 112. http://pralab.diee.unica.it @biggiobattista Experiments on MNIST digits Single-point attack • Linear SVM; 784 features; TR: 100; VAL: 500; TS: about 2000 – ‘0’ is the malicious (attacking) class – ‘4’ is the legitimate (attacked) one xc (0) xc 112[Biggio, Nelson, Laskov. Poisoning attacks against SVMs. ICML, 2012]
  113. 113. http://pralab.diee.unica.it @biggiobattista Experiments on MNIST digits Multiple-point attack • Linear SVM; 784 features; TR: 100; VAL: 500; TS: about 2000 – ‘0’ is the malicious (attacking) class – ‘4’ is the legitimate (attacked) one 113[Biggio, Nelson, Laskov. Poisoning attacks against SVMs. ICML, 2012]
  114. 114. http://pralab.diee.unica.it @biggiobattista How about Poisoning Deep Nets? • ICML 2017 Best Paper by Koh et al., “Understanding black-box predictions via Influence Functions” has derived adversarial training examples against a DNN – they have been constructed attacking only the last layer (KKT-based attack against logistic regression) and assuming the rest of the network to be ”frozen” 114
  115. 115. http://pralab.diee.unica.it @biggiobattista Towards Poisoning Deep Neural Networks • Solving the poisoning problem without exploiting KKT conditions (back-gradient) – Muñoz-González, Biggio, Roli et al., AISec 2017 https://arxiv.org/abs/1708.08689 115
  116. 116. 116 Countering Poisoning Attacks What is the rule? The rule is protect yourself at all times (from the movie “Million dollar baby”, 2004)
  117. 117. http://pralab.diee.unica.it @biggiobattista Security Measures against Poisoning • Rationale: poisoning injects outlying training samples • Two main strategies for countering this threat 1. Data sanitization: remove poisoning samples from training data • Bagging for fighting poisoning attacks • Reject-On-Negative-Impact (RONI) defense 2. Robust Learning: learning algorithms that are robust in the presence of poisoning samples xc (0) xc xc (0) xc 117
  118. 118. http://pralab.diee.unica.it @biggiobattista Robust Regression with TRIM • TRIM learns the model by retaining only training points with the smallest residuals argmin ',),* + ,, -, . = 1 |.| 2 3∈* 5 63 − 83 9 + ;Ω(>) @ = 1 + A B, . ⊂ 1, … , @ , . = B [Jagielski, Biggio et al., IEEE Symp. Security and Privacy, 2018] 118
  119. 119. http://pralab.diee.unica.it @biggiobattista Experiments with TRIM (Loan Dataset) • TRIM MSE is within 1% of original model MSE Existing methods Our defense No defense Better defense 119[Jagielski, Biggio et al., IEEE Symp. Security and Privacy, 2018]
  120. 120. 120 Other Attacks against ML
  121. 121. http://pralab.diee.unica.it @biggiobattista Attacks against Machine Learning Integrity Availability Privacy / Confidentiality Test data Evasion (a.k.a. adversarial examples) - Model extraction / stealing Model inversion (hill-climbing) Membership inference attacks Training data Poisoning (to allow subsequent intrusions) – e.g., backdoors or neural network trojans Poisoning (to maximize classification error) - [Biggio & Roli, Wild Patterns, 2018 https://arxiv.org/abs/1712.03141] Misclassifications that do not compromise normal system operation Misclassifications that compromise normal system operation Attacker’s Goal Attacker’s Capability Querying strategies that reveal confidential information on the learning model or its users Attacker’s Knowledge: • perfect-knowledge (PK) white-box attacks • limited-knowledge (LK) black-box attacks (transferability with surrogate/substitute learning models) 121
  122. 122. http://pralab.diee.unica.it @biggiobattista Model Inversion Attacks Privacy Attacks • Goal: to extract users’ sensitive information (e.g., face templates stored during user enrollment) – Fredrikson, Jha, Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. ACM CCS, 2015 • Also known as hill-climbing attacks in the biometric community – Adler. Vulnerabilities in biometric encryption systems. 5th Int’l Conf. AVBPA, 2005 – Galbally, McCool, Fierrez, Marcel, Ortega-Garcia. On the vulnerability of face verification systems to hill-climbing attacks. Patt. Rec., 2010 • How: by repeatedly querying the target system and adjusting the input sample to maximize its output score (e.g., a measure of the similarity of the input sample with the user templates) 122 Reconstructed Image Training Image
  123. 123. http://pralab.diee.unica.it @biggiobattista Membership Inference Attacks Privacy Attacks (Shokri et al., IEEE Symp. SP 2017) • Goal: to identify whether an input sample is part of the training set used to learn a deep neural network based on the observed prediction scores for each class 123
  124. 124. http://pralab.diee.unica.it @biggiobattista Training data (poisoned) Backdoored stop sign (labeled as speedlimit) Backdoor Attacks Poisoning Integrity Attacks 124 Backdoor / poisoning integrity attacks place mislabeled training points in a region of the feature space far from the rest of training data. The learning algorithm labels such region as desired, allowing for subsequent intrusions / misclassifications at test time Training data (no poisoning) T. Gu, B. Dolan-Gavitt, and S. Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. In NIPS Workshop on Machine Learning and Computer Security, 2017. X. Chen, C. Liu, B. Li, K. Lu, and D. Song. Targeted backdoor attacks on deep learning systems using data poisoning. ArXiv e-prints, 2017. M. Barreno, B. Nelson, R. Sears, A. D. Joseph, and J. D. Tygar. Can machine learning be secure? In Proc. ACM Symp. Information, Computer and Comm. Sec., ASIACCS ’06, pages 16–25, New York, NY, USA, 2006. ACM. M. Barreno, B. Nelson, A. Joseph, and J. Tygar. The security of machine learning. Machine Learning, 81:121–148, 2010. B. Biggio, B. Nelson, and P. Laskov. Poisoning attacks against support vector machines. In J. Langford and J. Pineau, editors, 29th Int’l Conf. on Machine Learning, pages 1807–1814. Omnipress, 2012. B. Biggio, G. Fumera, and F. Roli. Security evaluation of pattern classifiers under attack. IEEE Transactions on Knowledge and Data Engineering, 26(4):984–996, April 2014. H. Xiao, B. Biggio, G. Brown, G. Fumera, C. Eckert, and F. Roli. Is feature selection secure against training data poisoning? In F. Bach and D. Blei, editors, JMLR W&CP - Proc. 32nd Int’l Conf. Mach. Learning (ICML), volume 37, pages 1689–1698, 2015. L. Munoz-Gonzalez, B. Biggio, A. Demontis, A. Paudice, V. Wongrassamee, E. C. Lupu, and F. Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In 10th ACM Workshop on Artificial Intelligence and Security, AISec ’17, pp. 27–38, 2017. ACM. B. Biggio and F. Roli. Wild patterns: Ten years after the rise of adversarial machine learning. ArXiv e-prints, 2018. M. Jagielski, A. Oprea, B. Biggio, C. Liu, C. Nita-Rotaru, and B. Li. Manipulating machine learning: Poisoning attacks and countermeasures for regression learning. In 39th IEEE Symp. on Security and Privacy, 2018. Attackreferred toasbackdoor Attackreferredtoas ‘poisoningintegrity’
  125. 125. 125 Are Adversarial Examples a Real Security Threat?
  126. 126. http://pralab.diee.unica.it @biggiobattista World Is Not Digital… • ….Previous cases of adversarial examples have common characteristic: the adversary is able to precisely control the digital representation of the input to the machine learning tools….. 126 [M. Sharif et al., ACM CCS 2016] School Bus (x) Ostrich Struthio Camelus Adversarial Noise (r)
  127. 127. 127 Do Adversarial Examples Exist in the Physical World?
  128. 128. http://pralab.diee.unica.it @biggiobattista • Adversarial images fool deep networks even when they operate in the physical world, for example, images are taken from a cell-phone camera? – Alexey Kurakin et al. (2016, 2017) explored the possibility of creating adversarial images for machine learning systems which operate in the physical world. They used images taken from a cell-phone camera as an input to an Inception v3 image classification neural network – They showed that in such a set-up, a significant fraction of adversarial images crafted using the original network are misclassified even when fed to the classifier through the camera [Alexey Kurakin et al., ICLR 2017] Adversarial Images in the Physical World 128
  129. 129. http://pralab.diee.unica.it @biggiobattista [M. Sharif et al., ACM CCS 2016] The adversarial perturbation is applied only to the eyeglasses image region Adversarial Glasses 129
  130. 130. http://pralab.diee.unica.it @biggiobattista Should We Be Worried ? 130
  131. 131. http://pralab.diee.unica.it @biggiobattista No, We Should Not… 131 In this paper, we show experiments that suggest that a trained neural network classifies most of the pictures taken from different distances and angles of a perturbed image correctly. We believe this is because the adversarial property of the perturbation is sensitive to the scale at which the perturbed picture is viewed, so (for example) an autonomous car will misclassify a stop sign only from a small range of distances. [arXiv:1707.03501; CVPR 2017]
  132. 132. http://pralab.diee.unica.it @biggiobattista Yes, We Should... 132 [https://blog.openai.com/robust-adversarial-inputs/]
  133. 133. http://pralab.diee.unica.it @biggiobattista Yes, We Should... 133[Athalye et al., Synthesizing robust adversarial examples. ICLR, 2018]
  134. 134. http://pralab.diee.unica.it @biggiobattista Yes, We Should… 134
  135. 135. http://pralab.diee.unica.it @biggiobattista Adversarial Road Signs [Evtimov et al., CVPR 2017] 135
  136. 136. http://pralab.diee.unica.it @biggiobattista • Adversarial examples can exist in the physical world, we can fabricate concrete adversarial objects (glasses, road signs, etc.) • But the effectiveness of attacks carried out by adversarial objects is still to be investigated with large scale experiments in realistic security scenarios • Gilmer et al. (2018) have recently discussed the realism of security threat caused by adversarial examples, pointing out that it should be carefully investigated – Are indistinguishable adversarial examples a real security threat ? – For which real security scenarios adversarial examples are the best attack vector? Better than attacking components outside the machine learning component – … 136 Is This a Real Security Threat? [Justin Gilmer et al., Motivating the Rules of the Game for Adversarial Example Research, https://arxiv.org/abs/1807.06732]
  137. 137. 137 Are Indistinguishable Perturbations a Real Security Threat?
  138. 138. http://pralab.diee.unica.it @biggiobattista !(#) ≠ & The adversarial image x + r is visually hard to distinguish from x …There is a torrent of work that views increased robustness to restricted perturbations as making these models more secure. While not all of this work requires completely indistinguishable modications, many of the papers focus on specifically small modications, and the language in many suggests or implies that the degree of perceptibility of the perturbations is an important aspect of their security risk… Indistinguishable Adversarial Examples 138 [Justin Gilmer et al., Motivating the Rules of the Game for Adversarial Example Research, arXiv 2018]
  139. 139. http://pralab.diee.unica.it @biggiobattista • The attacker can benefit by minimal perturbation of a legitimate input; e.g., she could use the attack for a longer period of time before it is detected • But is minimal perturbation a necessary constraint for the attacker? Indistinguishable Adversarial Examples 139
  140. 140. http://pralab.diee.unica.it @biggiobattista • Is minimal perturbation a necessary constraint for the attacker? Indistinguishable Adversarial Examples 140
  141. 141. http://pralab.diee.unica.it @biggiobattista Attacks with Content Preservation 141 There are well known security applications where minimal perturbations and indistinguishability of adversarial inputs are not required at all…
  142. 142. http://pralab.diee.unica.it @biggiobattista …At the time of writing, we were unable to find a compelling example that required indistinguishability… To have the largest impact, we should both recast future adversarial example research as a contribution to core machine learning and develop new abstractions that capture realistic threat models. 142 Are Indistinguishable Perturbations a Real Security Threat? [Justin Gilmer et al., Motivating the Rules of the Game for Adversarial Example Research, arXiv 2018]
  143. 143. http://pralab.diee.unica.it @biggiobattista To Conclude… This is a recent research field… 143 Dagstuhl Perspectives Workshop on “Machine Learning in Computer Security” Schloss Dagstuhl, Germany, Sept. 9th-14th, 2012
  144. 144. http://pralab.diee.unica.it @biggiobattista Timeline of Learning Security 144 Security of DNNs Adversarial M L 2004-2005: pioneering work Dalvi et al., KDD 200; Lowd & M eek, KDD 2005 2013: Srndic & Laskov, NDSS claim nonlinear classifiers are secure 2006-2010: Barreno, Nelson, Rubinstein, Joseph, Tygar The Security of M achine Learning 2013-2014: Biggio et al., ECM L, IEEE TKDE high-confidence & black-box evasion attacks to show vulnerability of nonlinear classifiers 2014: Srndic & Laskov, IEEE S&P shows vulnerability of nonlinear classifiers with our ECM L ‘13 gradient-based attack 2014: Szegedy et al., ICLR adversarial exam ples vs DL 2016: Papernot et al., IEEE S&P evasion attacks / adversarial exam ples 2017: Papernot et al., ASIACCS black-box evasion attacks 2017: Carlini & W agner, IEEE S&P high-confidence evasion attacks 2017: Grosse et al., ESORICS Application to Android m alware 2017: Dem ontis et al., IEEE TDSC Secure learning for Android m alware 2004 2014 2006 2013 2014 2017 2016 2017 2017 2017 2014
  145. 145. http://pralab.diee.unica.it @biggiobattista Timeline of Learning Security 145 Security of DNNs Adversarial M L 2004-2005: pioneering work Dalvi et al., KDD 200; Lowd & M eek, KDD 2005 2013: Srndic & Laskov, NDSS claim nonlinear classifiers are secure 2006-2010: Barreno, Nelson, Rubinstein, Joseph, Tygar The Security of M achine Learning 2013-2014: Biggio et al., ECM L, IEEE TKDE high-confidence & black-box evasion attacks to show vulnerability of nonlinear classifiers 2014: Srndic & Laskov, IEEE S&P shows vulnerability of nonlinear classifiers with our ECM L ‘13 gradient-based attack 2014: Szegedy et al., ICLR adversarial exam ples vs DL 2016: Papernot et al., IEEE S&P evasion attacks / adversarial exam ples 2017: Papernot et al., ASIACCS black-box evasion attacks 2017: Carlini & W agner, IEEE S&P high-confidence evasion attacks 2017: Grosse et al., ESORICS Application to Android m alware 2004 2014 2006 2013 2014 2017 2016 2017 2017 2017 2014 2017: Dem ontis et al., IEEE TDSC Secure learning for Android m alware
  146. 146. http://pralab.diee.unica.it @biggiobattista Timeline of Learning Security AdversarialM L 2004-2005: pioneering work Dalvi et al., KDD 2004 Lowd & Meek, KDD 2005 2013: Srndic & Laskov, NDSS 2013: Biggio et al., ECML-PKDD - demonstrated vulnerability of nonlinear algorithms to gradient-based evasion attacks, also under limited knowledge Main contributions: 1. gradient-based adversarial perturbations (against SVMs and neural nets) 2. projected gradient descent / iterative attack (also on discrete features from malware data) transfer attack with surrogate/substitute model 3. maximum-confidence evasion (rather than minimum-distance evasion) Main contributions: - minimum-distance evasion of linear classifiers - notion of adversary-aware classifiers 2006-2010: Barreno, Nelson, Rubinstein, Joseph, Tygar The Security of Machine Learning (and references therein) Main contributions: - first consolidated view of the adversarial ML problem - attack taxonomy - exemplary attacks against some learning algorithms 2014: Szegedy et al., ICLR Independent discovery of (gradient- based) minimum-distance adversarial examples against deep nets; earlier implementation of adversarial training SecurityofDNNs 2016: Papernot et al., IEEE S&P Framework for security evalution of deep nets 2017: Papernot et al., ASIACCS Black-box evasion attacks with substitute models (breaks distillation with transfer attacks on a smoother surrogate classifier) 2017: Carlini & Wagner, IEEE S&P Breaks again distillation with maximum-confidence evasion attacks (rather than using minimum-distance adversarial examples) 2016: Papernot et al., Euro S&P Distillation defense (gradient masking) Main contributions: - evasion of linear PDF malware detectors - claims nonlinear classifiers can be more secure 2014: Biggio et al., IEEE TKDE Main contributions: - framework for security evaluation of learning algorithms - attacker’s model in terms of goal, knowledge, capability 2017: Demontis et al., IEEE TDSC Yes, Machine Learning Can Be More Secure! A Case Study on Android Malware Detection Main contributions: - Secure SVM against adversarial examples in malware detection 2017: Grosse et al., ESORICS Adversarial examples for malware detection 2018: Madry et al., ICLR Improves the basic iterative attack from Kurakin et al. by adding noise before running the attack; first successful use of adversarial training to generalize across many attack algorithms 2014: Srndic & Laskov, IEEE S&P used Biggio et al.’s ECML-PKDD ‘13 gradient-based evasion attack to demonstrate vulnerability of nonlinear PDF malware detectors 2006: Globerson & Roweis, ICML 2009: Kolcz et al., CEAS 2010: Biggio et al., IJMLC Main contributions: - evasion attacks against linear classifiers in spam filtering Work on security evaluation of learning algorithms Work on evasion attacks (a.k.a. adversarial examples) Pioneering work on adversarial machine learning ... in malware detection (PDF / Android) Legend 1 2 3 4 1 2 3 4 2015: Goodfellow et al., ICLR Maximin formulation of adversarial training, with adversarial examples generated iteratively in the inner loop 2016: Kurakin et al. Basic iterative attack with projected gradient to generate adversarial examples 2 iterative attacks Biggio and Roli, Wild Patterns: Ten Years After The Rise of Adversarial Machine Learning, Pattern Recognition, 2018 146
  147. 147. http://pralab.diee.unica.it @biggiobattista Black Swans to the Fore 147 [Szegedy et al., Intriguing properties of neural networks, 2014] After this “black swan”, the issue of security of DNNs came to the fore… Not only on scientific specialistic journals…
  148. 148. http://pralab.diee.unica.it @biggiobattista The Safety Issue to the Fore… 148 The black box of AI D. Castelvecchi, Nature, Vol. 538, 20, Oct 2016 Machine learning is becoming ubiquitous in basic research as well as in industry. But for scientists to trust it, they first need to understand what the machines are doing. Ellie Dobson, director of data science at the big-data firm Arundo Analytics in Oslo: If something were to go wrong as a result of setting the UK interest rates, she says, “the Bank of England can’t say, the black box made me do it’”.
  149. 149. http://pralab.diee.unica.it @biggiobattista Why So Much Interest? 149 Before the deep net “revolution”, people were not surprised when machine learning was wrong, they were more amazed when it worked well… Now that it seems to work for real applications, people are disappointed, and worried, for errors that humans do not do…
  150. 150. http://pralab.diee.unica.it @biggiobattista Errors of Humans and Machines… 150 Machine learning decisions are affected by several sources of bias that causes “strange” errors But we should keep in mind that also humans are biased…
  151. 151. http://pralab.diee.unica.it @biggiobattista The Bat and the Ball Problem 151 A bat and a ball together cost $ 1.10 The bat costs $ 1.0 more than the ball How much does the ball cost ? Please, give me the first answer coming to your mind !
  152. 152. http://pralab.diee.unica.it @biggiobattista The Bat and the Ball Problem 152 Exact solution is 0.05 dollar (5 cents) The wrong solution ($ 0.10) is due to the attribute substitution, a psychological process thought to underlie a number of cognitive biases It occurs when an individual has to make a judgment (of a target attribute) that is computationally complex, and instead substitutes a more easily calculated heuristic attribute bat+ball=$1.10 bat=ball+$1.0 ⎧ ⎨ ⎪ ⎩⎪
  153. 153. http://pralab.diee.unica.it @biggiobattista Trust in Humans or Machines? Algorithms are biased, but also humans are as well… When should you trust humans and when algorithms? 153
  154. 154. http://pralab.diee.unica.it @biggiobattista Learning Comes at a Price! 154 The introduction of novel learning functionalities increases the attack surface of computer systems and produces new vulnerabilities Safety of machine learning will be more and more important in future computer systems, as well as accountability, transparency, and the protection of fundamental human values and rights
  155. 155. http://pralab.diee.unica.it @biggiobattista Thanks for Listening! Any questions? 155 Engineering isn't about perfect solutions; it's about doing the best you can with limited resources (Randy Pausch, 1960-2008)

×