Machine Learning and Artificial Intelligence Security Risk: Categorizing Attacks and Failure Modes

Learn how and why machine learning and artificial intelligence technology fails and understand ways to make these systems more secure and resilient.

Introduction

• Machine learning security concerns
• What you should know

1. Machine Learning Foundations

• How systems can fail and how to protect them
• Why does ML security matter
• Attacks vs. unintentional failure modes
• Security goals for ML: CIA

2. Intentional Failure Modes/Attacks

• Perturbation attacks and AUPs
• Poisoning attacks
• Reprogramming neural nets
• Physical domain (3D adversarial objects)
• Supply chain attacks
• Model inversion
• System manipulation
• Membership inference and model stealing
• Backdoors and existing exploits

3. Unintentional Failure Modes/Intrinsic Design Flaws

• Reward hacking
• Side effects in reinforcement learning
• Distributional shifts and incomplete testing
• Overfitting/underfitting
• Data bias considerations

4. Building Resilient ML

• Effective techniques for building resilience in ML
• ML dataset hygiene
• ML adversarial training
• ML access control to APIs

Conclusion

• Next steps