Security Risks in AI and Machine Learning: Categorizing Attacks and Failure Modes

Contents

• Introduction Introduction

  • Machine learning security concerns

    1m 21s
  • What you should know

    25s
• 1. Machine Learning Foundations 1. Machine Learning Foundations

  • How systems can fail and how to protect them

    3m 22s
  • Why does ML security matter

    5m 41s
  • Attacks vs. unintentional failure modes

    2m 59s
  • Security goals for ML: CIA

    2m 45s
• 2. Intentional Failure Modes/Attacks 2. Intentional Failure Modes/Attacks

  • Perturbation attacks and AUPs

    3m 31s
  • Poisoning attacks

    3m 11s
  • Reprogramming neural nets

    1m 39s
  • Physical domain (3D adversarial objects)

    2m 34s
  • Supply chain attacks

    2m 42s
  • Model inversion

    3m 12s
  • System manipulation

    3m 2s
  • Membership inference and model stealing

    2m 3s
  • Backdoors and existing exploits

    2m 19s
• 3. Unintentional Failure Modes/Intrinsic Design Flaws 3. Unintentional Failure Modes/Intrinsic Design Flaws

  • Reward hacking

    2m 16s
  • Side effects in reinforcement learning

    2m 30s
  • Distributional shifts and incomplete testing

    3m 1s
  • Overfitting/underfitting

    2m 45s
  • Data bias considerations

    4m 48s
• 4. Building Resilient ML 4. Building Resilient ML

  • Effective techniques for building resilience in ML

    2m 33s
  • ML dataset hygiene

    4m 26s
  • ML adversarial training

    4m 2s
  • ML access control to APIs

    2m 56s
• Conclusion Conclusion

  • Next steps

    1m 32s