AI & ML in Cyber Security - Why Algorithms Are Dangerous

Overview

Explore the limitations and potential dangers of machine learning and artificial intelligence in cybersecurity in this Black Hat conference talk. Delve into the issues of explainability in algorithms and learn where deep learning applications may be inappropriate. Examine real-world examples demonstrating how blind application of algorithms, including deep learning, can lead to incorrect results. Discover the importance of understanding data, choosing appropriate algorithms, and incorporating expert knowledge in cybersecurity applications. Gain insights into alternative approaches like probabilistic inference and belief networks for more robust security solutions.

Syllabus

Intro
RAFFAEL MARTY
OUTLINE
ML AND AI - WHAT IS IT? MACHINE LEARNING Algorithmic ways to describe data Supervised
MACHINE LEARNING USES IN SECURITY
FAMOUS AI (ALGORITHM) FAILURES
WHAT MAKES ALGORITHMS DANGEROUS? ALGORITHMS MAKE ASSUMPTIONS ABOUT THE DATA
COGNITIVE BIASES
THE DANGERS WITH DEEP LEARNING - WHEN NOT TO USE IT
ADVERSARIAL MACHINE LEARNING
DEEP LEARNING - THE SOLUTION TO EVERYTHING
UNSUPERVISED TO THE RESCUE?
UNDERSTAND AND CLEAN THE DATA
ENGINEERING DISTANCE FUNCTIONS
CHOOSING THE RIGHT UNSUPERVISED ALGORITHM
CHOOSING THE CORRECT ALGORITHM PARAMETERS
INTERPRETING THE DATA
A DIFFERENT APPROACH - PROBABILISTIC INFERENCE Rather than running algorithms the model the shape of data, we need to take expert knowledge/ domain expertise into account
ST STEP-BUILD THE GRAPH
ND STEP - GROUP NODES
RD STEP - INTRODUCE DEPENDENCIES
TH STEP - ESTIMATE PROBABILITIES
TH STEP-GOAL COMPUTATION
TH STEP-OBSERVE ACTIVITIES
TH STEP-EXPERT INPUT Strengthen the network by introducing expert knowledge
BELIEF NETWORKS - SOME OBSERVATIONS