Course Objective & Scope
Proprietary machine learning models are targeted not only through infrastructure flaws, but also through prediction APIs—allowing attackers to clone a model via query-based extraction, infer whether specific records were in the training set (membership inference), reconstruct sensitive inputs (model inversion) or manipulate predictions. This course provides an end-to-end view of ML model security across the model lifecycle using the MITRE ATLAS framework.
Participants execute foundational privacy and integrity attacks against a live scoring model and leave with a reusable enterprise model security checklist covering API hardening, output rounding, query anomaly detection and model watermarking.
Who Should Attend?
- Data scientists, machine learning engineers and AI engineers
- Backend and platform teams exposing model prediction APIs
- Application security (AppSec) and cybersecurity specialists
- Prerequisite: foundational Python and machine learning (scikit-learn / PyTorch) knowledge
Key Takeaways
- Map confidentiality, integrity and availability risks of an ML system using the MITRE ATLAS matrix
- Simulate and detect query-based model extraction (model stealing) attacks
- Quantify training data leakage risk through membership inference and model inversion tests
- Harden prediction APIs with confidence-score rounding, query entropy monitoring and rate quotas
- Compile a reusable pre-production model security verification checklist
Curriculum
The ML Threat Landscape and the MITRE ATLAS Framework
- Confidentiality, integrity and availability across data, training, model artefacts and inference APIs
- White-box, grey-box and black-box threat actor assumptions
Model Extraction and Intellectual Property Protection
- Training surrogate models and cloning decision boundaries via prediction API queries
- Logit suppression, PRADA-style query distribution monitoring and model watermarking
Model Privacy: Membership Inference and Model Inversion
- How overfitting enables training data leakage and shadow-model membership attacks
- Mitigations via regularisation, differential privacy and output precision coarsening
Overview of Integrity and Availability Threats
- Evasion, data poisoning and resource-exhausting sponge/latency inputs
- Model serialisation security (pickle vs safetensors/ONNX) and artefact integrity checks
Hands-On Model Security Audit
- Running extraction and privacy audit tests against a sample credit/risk scoring model
- Producing a reusable model security checklist and defence report for internal teams
Other Courses in This Category
-
Hands-on training on FGSM, PGD, Carlini & Wagner, AutoAttack and black-box evasion attacks using the IBM Adversarial Robustness Toolbox (ART), paired with adversarial training and certified robustness defenses.
- Live Course
- 5 modules
-
Teaches how to detect and prevent label flipping, clean-label poisoning and hidden backdoor/trojan attacks in training and fine-tuning datasets using spectral signatures, activation clustering and data sanitisation.
- Live Course
- 5 modules
-
In-depth training on protecting vision, audio, biometric and multimodal deep neural networks against weight and architecture theft, feature reconstruction, sponge latency attacks and inference-layer compromises.
- Live Course
- 5 modules
-
Teaches how to secure the end-to-end MLOps lifecycle—across data pipelines, Jupyter workspaces, MLflow, Kubeflow, model registries, CI/CD and KServe/Kubernetes serving—using Sigstore/Cosign, ModelScan and Zero Trust controls.
- Live Course
- 5 modules
-
Covers the full OWASP Top 10 for LLM Applications—including direct and indirect prompt injection, improper output handling, unbounded consumption and sensitive data disclosure—from a secure software development (SSDLC) perspective.
- Live Course
- 5 modules
-
Hands-on training on conducting end-to-end penetration tests against classical ML, deep learning and LLM APIs using MITRE ATLAS, Counterfit, ART, garak and PyRIT, and producing actionable pentest reports.
- Live Course
- 5 modules
Get in touch
