Course Objective & Scope
Machine learning models used in fraud detection, credit scoring, computer vision and anomaly detection are vulnerable to training data poisoning, adversarial examples that cross decision boundaries and queries that extract model weights or sensitive training records. Guided by MITRE ATLAS and NIST AI 100-2, this advanced programme examines offensive and defensive techniques across the ML lifecycle.
In a Python lab environment (using tools such as the Adversarial Robustness Toolbox), participants test evasion, poisoning and privacy attacks against sample models and implement adversarial training, data sanitization and runtime monitoring controls.
Who Should Attend?
- Data scientists, machine learning and AI engineers
- AI security specialists and Red Team practitioners
- MLOps engineers deploying high-stakes decision models to production
- Prerequisite: Python, core ML/deep learning concepts and hands-on experience with PyTorch or scikit-learn
Key Takeaways
- Classify threats across ML pipelines using the MITRE ATLAS and NIST AI 100-2 taxonomies
- Evaluate model robustness with white-box and black-box evasion attacks (FGSM, PGD, query-based)
- Detect and filter data poisoning and backdoor/trojan triggers in training datasets
- Assess model extraction (model stealing) and membership inference privacy risks
- Harden models through adversarial training, input validation and runtime anomaly detection
Curriculum
Adversarial ML Taxonomy and MITRE ATLAS
- Training-time versus inference-time attack surfaces
- Attacker knowledge models (white-box, gray-box, black-box) and MITRE ATLAS case studies
Evasion Attacks and Crafting Adversarial Examples
- Gradient-based (FGSM, PGD) and query-based black-box evasion attacks
- Constrained perturbations on tabular models (fraud/credit) as well as vision and text models
Data Poisoning and Backdoor/Trojan Attacks
- Label flipping, clean-label poisoning and hidden backdoor triggers
- Supply chain risks in public datasets and pre-trained model weights
Model Extraction and Privacy Attacks
- Model stealing via prediction APIs using query synthesis and distillation
- Measuring training data leakage with membership inference and model inversion
Defenses, Robustness Testing and Runtime Monitoring
- Adversarial training, data sanitization, differential privacy and API query throttling
- Building a robustness evaluation report and checklist with the Adversarial Robustness Toolbox (ART)
Other Courses in This Category
-
OWASP Top 10 for LLM & Agentic Applications, direct/indirect prompt injection, RAG data leakage, MCP/tool-calling authorization and guardrail architectures for engineers building secure AI systems.
- Live Course
- 5 modules
-
ISO/IEC 27001:2022 clauses, asset and risk assessment, Annex A controls, Statement of Applicability (SoA) and internal audit readiness for teams establishing and operating an ISMS.
- Live Course
- 5 modules
-
The Linux network stack, nftables/iptables, SSH and privilege hardening, SELinux/AppArmor, auditd and packet analysis for systems and security engineers securing production Linux servers.
- Live Course
- 5 modules
-
Static and dynamic malware analysis in isolated labs, PE/ELF inspection, behavioural monitoring, unpacking, code analysis with Ghidra and YARA rule authoring for security analysts.
- Live Course
- 5 modules
-
All 8 domains of the (ISC)² CISSP Common Body of Knowledge (CBK), risk and governance mindset, security architecture and scenario-based question analysis for experienced security professionals.
- Live Course
- 5 modules
-
The CIA triad, modern threat vectors, identity and access management, network security and baseline hardening practices for IT and software teams who need to identify and mitigate enterprise security risks.
- Live Course
- 5 modules
Get in touch
