Slide 1: Adversarial ML: How Artificial Intelligence is Enabling Cyber Resilience
Adversarial ML: How Artificial Intelligence is Enabling Cyber Resilience Michael Slawinski, Ph.D. Josh Fu, CISM, CISSP Staff Data Scientist Security Engineer BlackBerry Cylance BlackBerry Cylance
Slide 2: Agenda
Agenda 1. File Classification via Machine Learning 2. Weaknesses of Machine Learning Classifiers 3. How these weaknesses can be exploited to make a malicious file look benign 4. How to harden your model against adversarial attacks
Slide 3: Classification
Classification • How to tell if a script/executable/word document/PDF is malicious? • Many security vendors have large labeled datasets Malicious • Use these labeled samples to train a classifier • How can this classifier be attacked? • Is it possible to perturb a file in such a way as to cross the
Slide 4: Naiveté is Dangerous
Naiveté is Dangerous We make the mistake of assuming the model is judging as we judge. In other words, we assume the machine learning model has baked into it a conceptual understanding of the objects being classified… Example: Lie Detectors – what is a lie? Human Point of View Lie Detector Point of
Slide 5: Machine Learning – definition and utility
Machine Learning – definition and utility Definition: The design and implementation of learning algorithms. Learning Algorithms: Algorithms which are not explicitly programmed Decision making scheme is a result of optimizing some objective function based on observed data Utility of Learning Algorith
Slide 6: Machine Learning – definition and utility
Machine Learning – definition and utility
Slide 7: Adversarial Examples
Adversarial Examples Model says ‘malicious’ Model says ‘benign’ If a new sample lands within the Functionally Model point of view boundary, we classify it as benign. Otherwise, malicious. = ≠
Slide 8: Adversarial Machine Learning
Adversarial Machine Learning gibbon panda Decision boundary benign.exe malware.exe Decision boundary
Slide 9: Vulnerability to Adversarial Attacks – 'Snow Features'
Vulnerability to Adversarial Attacks – ‘Snow Features’ Training Set Model Says: ‘Wolf’ 99.9% Upshot: Model overemphasizing white, i.e., snow pixels due to high wolf/snow correlation
Slide 10: Vulnerability to Adversarial Attacks
Vulnerability to Adversarial Attacks Snow Features: model features which are not fundamental to the nature of the object being classified, but which aid in classification due to co-occurrence. Models trained on such features are vulnerable to adversarial attacks Key Takeaway: Malicious and Benign fi
Slide 11: CylancePROTECT Defenses Against Adversarial Attacks
CylancePROTECT® Defenses Against Adversarial Attacks 1. Reduce leveraging of ‘snow features’ • Features which when perturbed, change the nature of the original file 2. Adversarial Training • Leverage more sophisticated learning algorithms 3. Anomaly Detection • Learn probability distribution of the
Slide 12: Practical Applications
Practical Applications BlackBerry. | YL /\NCE.
Slide 13: Why CAPTCHAs
Why CAPTCHAs • CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) • Turing Test Source: https://www.maketecheasier.com/captchas-why-we-need-them/
Slide 14: Usage vs Trickery
Usage vs Trickery • CNN for CAPTCHAs • Natural Language Processing (NLP) ▪Twitter Bot - John Seymour and Philip Tully at ZeroFOX in 2016 ▪Laser phishing • Speech to Text
Slide 15: Today's CAPTCHA Breaking Techniques
Today’s CAPTCHA Breaking Techniques I will always choose a lazy person to do a difficult job…because he will find an easy way to do it. - Bill Gates
Slide 16: Browser Extension
Browser Extension • Buster: Captcha Solver for Humans by Armin Sebastian • Uses Google’s own Natural Language Processing • This is the state of security today
Slide 17: Noise as Defense
Noise as Defense ▪ Mitigate model discovery ▪ Optical vs Digital classification ▪ Noise is one way of enabling resilience TensorFlow Tutorial #11 Adversarial Examples
Slide 18: Key Takeaways
Key Takeaways 1. Tools can be easy and cheap 2. ML will have a role in the arms race 3. The next frontier for ML → solving identity and authentication challenges
Slide 19: How ML Applies to Cylance
How ML Applies to Cylance • Recognition of files → CylancePROTECT and Cylance Smart Antivirus™ • Recognition of processes → CylanceOPTICS™ • Recognition of people → CylancePERSONA • Additional Benefits: ▪ Advanced enough to even work offline and on old systems with no updates ▪ Faster threat hunting
Slide 20: Get a Demo
Get a Demo See the next generation in endpoint security solutions – and take that first step towards getting your organization to a state of prevention.
Slide 21: THANK YOU!
THANK YOU! • Cylance www.cylance.com • Best Machine Learning Resources for Starters https://machinelearningmastery.com/best-machine-learning-resources-for-getting-started/ • Twitter Story https://www.nytimes.com/2018/11/19/science/artificial-intelligence-deepfakes-fake-news.html • Google cloud Visio
Slide 22: Questions and Answers
Questions and Answers BlackBerry. | YA NCE.
Slide 23: ©2019 Cylance Inc. All Rights Reserved.
©2019 Cylance Inc. All Rights Reserved.