Inhalt
| Inhalt |
The lecture Machine Learning Security (MLS) covers multiple kinds of relations between machine learning (ML) and IT security. After a recap of the fundamentals of both of these fields, we discuss three roles of ML in IT security: ML as defense, ML as victim, and ML as attack tool. Finally, we will look at methods to securing ML. |
| Literatur |
Announced in the lecture |
| Lernziele |
At completion of the course, you will:
- understand existing threats to ML as well as possible countermeasures,
- understand the application of ML in security,
- be able to implement robust and secure machine-learning systems,
- have developed practical skills in using ML-based tools for solving real-world problems in security,
- be able to evaluate ML models for their security robustness.
|
| Voraussetzungen |
Formal requirements: None
Recommended knowledge from or equivalent to the lectures:
- Datensicherheit
- Systemsicherheit
- Künstliche Intelligenz
- Deep Learning
|
| Leistungsnachweis |
Oral exam
Grade bonus (0.3) for successful lab participation |
| Lerninhalte |
Upon successful completion of this module, students will be able to:
- Understand the comprehensive taxonomy of threats to machine learning systems, specifically regarding how adversaries exploit the statistical nature of AI to compromise integrity, availability, and privacy.
- Identify and analyze diverse attack vectors used against predictive and generative AI, including training-stage poisoning, deployment-stage evasion attacks, and Large Language Model (LLM) vulnerabilities like prompt injection.
- Apply machine learning as a powerful defensive tool (#secBYai) to enhance IT security through tasks such as network intrusion detection, malware behavior analysis, and automated spam or phishing detection.
- Implement robust and secure machine learning systems by developing practical skills to solve real-world security problems with appropriate algorithms and tools.
- Critically evaluate ML models and workflows while identifying and mitigating common pitfalls such as sampling bias, label inaccuracy, and the base rate fallacy.
- Navigate the complexities of LLM security, including the implementation of alignment techniques, input/output filtering, and prompt augmentation to increase jailbreak resilience.
- Perform security assurance and behavior auditing to inspect model decisions for unintended logic or vulnerabilities.
- Utilize advanced evaluation methodologies, such as adversarial testing and the "LLM as a Judge" framework, to conduct automated and consistent assessments of AI system robustness.
|
