AI Security
Defend ML systems and LLM applications against adversarial attacks, pipeline abuse, prompt injection, model theft, and supply-chain compromise. Machine learning and generative AI expand the attack surface for software systems.
Level: Advanced · Category: Cybersecurity · Estimated time: 7 hours
Prerequisites
- Machine Learning Basics
Lessons
- Adversarial ML Fundamentals — Threat model for ML, adversarial examples, robustness vs accuracy, and where attacks surface in deployment.
- Evasion Attacks & Defenses — FGSM, PGD, adversarial training, input sanitization, and detection — defending inference-time attacks.
- Data Poisoning & Backdoor Attacks — Poisoning training data, backdoored models, dataset provenance, and validation strategies.
- Securing ML Pipelines — Secrets management, least privilege, artifact signing, reproducible builds, monitoring, and safe CI/CD for models.
- Prompt Injection & LLM Security — Direct and indirect prompt injection, tool abuse, separation of instructions and data, guardrails, and secure agent design.
- Model Extraction & Privacy Attacks — Query-based model stealing, distillation-style extraction, membership inference, and API rate-limiting patterns.
- ML Supply Chain Security — Dependency risk, malicious packages, Hugging Face and checkpoint provenance, signing, and vetting third-party models.
Related consolidated topics
- AI Security & Adversarial ML: Understand and defend against adversarial attacks on ML systems.
Topics covered
ai-security, adversarial-ml, prompt-injection, ml-pipeline-security, model-extraction, supply-chain, llm-security, robustness