JARVIS Never Saw It Coming - Hacking Machine Learning in Speech, Text and Face Recognition

Explore the world of hacking machine learning systems in this conference talk from 44CON 2018. Delve into the emerging field of Adversarial ML, learning how to exploit weak points in speech, text, and face recognition algorithms. Discover techniques for achieving unexpected consequences, data leakage, memory corruption, and output manipulation in ML systems. Witness a live demonstration showcasing the potential vulnerabilities in these intelligent systems. Gain insights into the top 5 attacks based on CVSS and business impact, and understand where to focus your offensive research. Learn about various attack methods, including cloning, backdoors, encoding, and adversarial attacks on audio and visual recognition systems. Examine real-world examples, such as misclassifying rifles as bananas and evading next-generation antivirus software using AI. Equip yourself with knowledge to better understand and address the security challenges posed by machine learning technologies.

Intro
HOW DID WE GET HERE?
CLEVER HANS
ARTIFICIAL INTELLIGENCE?
INTELLIGENT SYSTEM
WHAT IS A ML MODEL?
CODE POINT OF VIEW
FROM TRAINING TO INFERENCE
BIAS - SOLVING THE WRONG PROBLEM
TOP 5 ATTACKS (CVSS)
TOP 5 ATTACKS (BUSINESS IMPACT)
WHERE TO ATTACK?
PRELIMINARY RESULTS
ATTACK OF THE CLONES
BACKDOORS
ENCODING
MISS-PREDICTIONS (ADVERSARIAL ATTACKS)
TURTLE OR A RIFLE?
ADVERSARIAL AUDIO
EVADING NEXT GENERATION AV USING AI
ACKNOWLEDGMENTS
REFERENCES