Industrial practitioners' mental models of adversarial machine learning
summary
Qualitative study of 15 interviews with European startup practitioners in 2020 uses semi-structured interviews plus a whiteboard drawing task and thematic coding, finding frequent confusion between adversarial ML and general cybersecurity, that only half mention ML-specific mitigations, and that security is framed around whole workflows not isolated models. Pipelines are shown as complex systems with feedback and multiple components, one reported real-world case resembles semi‑automated fraud, implying need to integrate AML into practices and targeted education.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:12Outline
- 2:12Adversarial Machine Learning
- 3:29Qualitative Sample - 15 Participants (2020)
- 4:22Interview Procedure
- 5:18Key findings - AML versus Non-AML Security
- 6:06Details - AML versus Non-AML Security
- 9:10Key findings - Model versus Workflows
- 9:50Details - Model versus Workflows
- 10:45Open questions
- 11:33AML attacks in practice
- 12:14Implications
related talks
53mSecurity of Machine Learning Systems
44mAdvancing Cybersecurity with AI and Machine Learning: Opportunities and Challenges
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
36mAdversarial Machine Learning: How Secure Is ML?
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
32mAI/Machine Learning: What is Actually Working in Cybersecurity?
01h 05mNavigating the Security and Privacy Landscape of Modern AI
50mBuilding AI Security In: MLSecOps in Practice