Course contentsShow
← Security
Lesson 2872 of 3,05253. AI and ML SecurityFree lesson

ML Supply Chain Threat Landscape

Understanding attack vectors across data collection, model training, distribution, and deployment pipelines.

ML Supply Chain Threat Landscape

What you'll learn: You'll explore how attackers can compromise machine learning systems at every stage—from raw data gathering through model deployment—by targeting the ML development pipeline itself.

Understanding the Attack Surface

Machine learning systems don't emerge fully formed. They're built through a supply chain that includes data collection, labeling, model training, packaging, distribution, and deployment. Each stage represents a potential attack vector.

Think of it like manufacturing a car: attackers could sabotage the steel supplier, the assembly line, the shipping container, or the dealership. In ML, the "product" is a trained model, and adversaries can strike at any point along its journey.

Key Attack Vectors Across the Pipeline

Data Collection Stage

Attackers can inject malicious samples into training datasets through compromised data sources, scraped websites, or public repositories. This enables data poisoning attacks you've already studied—but here the threat comes from upstream suppliers or partners.

Model Training Stage

Compromised training infrastructure (cloud services, GPUs, framework libraries) can introduce backdoors during the training process itself. Malicious dependencies in PyTorch, TensorFlow, or custom libraries can alter model behavior without touching the data.

Distribution Stage

Pre-trained models hosted on model hubs (like Hugging Face, TensorFlow Hub) can be trojaned or replaced with malicious versions. Developers downloading "ResNet-50" might unknowingly get a backdoored variant.

Deployment Stage

Even secure models can be compromised through malicious inference servers, corrupted model serialization formats (pickle files), or supply chain attacks on deployment containers and MLOps tools.

Why This Matters

Unlike traditional software, ML systems inherit trust from external artifacts—datasets, pre-trained weights, labeling services. A single compromised component can invisibly corrupt the entire system's behavior.

Key Takeaway: The ML supply chain spans data sourcing, training infrastructure, model repositories, and deployment pipelines—each representing distinct attack surfaces that adversaries can exploit to compromise model integrity, inject backdoors, or steal intellectual property.