Securing AI Models: Threats, Attacks, and Defenses

• An understanding of machine learning workflows and model training
• Experience with Python and common ML frameworks such as PyTorch or TensorFlow
• Familiarity with basic security or threat modeling concepts is helpful

Audience

• Machine learning engineers
• Cybersecurity analysts
• AI researchers and model validation teams

Securing AI Models is the discipline of defending machine learning systems against model-specific threats such as adversarial inputs, data poisoning, inversion attacks, and privacy leakage.

This instructor-led, live training (online or onsite) is aimed at intermediate-level machine learning and cybersecurity professionals who wish to understand and mitigate emerging threats against AI models, using both conceptual frameworks and hands-on defenses like robust training and differential privacy.

By the end of this training, participants will be able to:

• Identify and classify AI-specific threats such as adversarial attacks, inversion, and poisoning.
• Use tools like the Adversarial Robustness Toolbox (ART) to simulate attacks and test models.
• Apply practical defenses including adversarial training, noise injection, and privacy-preserving techniques.
• Design threat-aware model evaluation strategies in production environments.

Format of the Course

• Interactive lecture and discussion.
• Lots of exercises and practice.
• Hands-on implementation in a live-lab environment.

Course Customization Options

• To request a customized training for this course, please contact us to arrange.

Introduction to AI Threat Modeling

• What makes AI systems vulnerable?
• AI attack surface vs traditional systems
• Key attack vectors: data, model, output, and interface layers

Adversarial Attacks on AI Models

• Understanding adversarial examples and perturbation techniques
• White-box vs black-box attacks
• FGSM, PGD, and DeepFool methods
• Visualizing and crafting adversarial samples

Model Inversion and Privacy Leakage

• Inferring training data from model output
• Membership inference attacks
• Privacy risks in classification and generative models

Data Poisoning and Backdoor Injections

• How poisoned data influences model behavior
• Trigger-based backdoors and Trojan attacks
• Detection and sanitization strategies

Robustness and Defense Techniques

• Adversarial training and data augmentation
• Gradient masking and input preprocessing
• Model smoothing and regularization techniques

Privacy-Preserving AI Defenses

• Introduction to differential privacy
• Noise injection and privacy budgets
• Federated learning and secure aggregation

AI Security in Practice

• Threat-aware model evaluation and deployment
• Using ART (Adversarial Robustness Toolbox) in applied settings
• Industry case studies: real-world breaches and mitigations

Summary and Next Steps