Async review
Instruction and application
In ProgressRecap core topics:
- Unit 1: Security Fundamentals in Machine Learning
Unit 1: Security Fundamentals in Machine Learning
In Unit 1, you explored…
Core security principles
– Confidentiality, integrity, authentication, non-repudiation, and service integrity in the context of ML workflows.
Lifecycle threats
– Common vulnerabilities across the ML lifecycle, including data poisoning, model inversion, insecure APIs, and audit gaps.
Infrastructure decisions
– Trade-offs between local, cloud, and hybrid setups in terms of control, compliance, scalability, and risk exposure.
Secure workflow design
– Techniques such as threat modelling, least privilege, encryption, and monitoring embedded into ML pipelines.
Team protocols and culture
– Role-based responsibilities, incident response practices, and a culture of continuous security improvement.
Five principles shape every stage of a secure ML system
PrincipleFocus areaML exampleConfidentialityPreventing unauthorised access to data/modelsEncrypting training data and protecting model outputsIntegrityEnsuring correctness of data and models Detecting tampering in training labels or model weightsAuthenticationVerifying user/system identity Using API keys and MFA to restrict model accessNon-repudiation Enabling traceability of actions Logging model updates and access eventsService integrity Defending runtime model behavior Blocking adversarial inputs and monitoring for drift
Threats across the ML lifecycle: Where can things go wrong?
Data poisoning:
Malicious input during training corrupts model behaviour.
Model inversion:
Attackers reconstruct training data from outputs.
Adversarial inputs:
Carefully crafted inputs trick deployed models.
API misuse:
Open or unauthenticated endpoints expose model logic or data.Secure ML practices
- Encrypt sensitive assets.
- Apply role-based access controls.
- Use secure APIs with rate limiting.
- Log access and model events.
Who protects what in your ML system?
Security in ML systems is a shared responsibility across roles:
Data scientists:
Monitor model behaviour and validate training data.
Engineers:
Secure data pipelines and configure access controls.
Platform admins:
Manage infrastructure and enforce CI/CD security.
Action item: Quick reflection - What matters most for your role?
Which security concept feels most critical for your role—and why?
Share your response in the chat!