Supply Chain Poisoning in ML Pipelines
What you'll learn: How attackers compromise the ML supply chain—datasets, pre-trained models, or training infrastructure—to inject backdoors before models ever reach production.
The Attack Vector
While the previous lessons covered poisoning attacks that target your own training process, supply chain poisoning strikes earlier: attackers compromise upstream dependencies. Instead of attacking your pipeline directly, they poison components you download and trust—pre-trained models from public repositories, curated datasets from third parties, or even the infrastructure services you use for training.
Think of it like buying pre-made cookie dough from the store. You trust the manufacturer, but if someone poisoned the dough at the factory, every batch you bake is compromised—and you never see it coming.
How It Works
Attackers target three main vectors:
Pre-trained Model Poisoning: Popular model hubs host millions of models. An attacker uploads a "state-of-the-art" vision model with a backdoor already embedded. When you fine-tune it for your task, the trigger persists.
Dataset Compromise: Poisoning widely-used training datasets (scraped web data, benchmark sets) means hundreds of downstream models inherit the backdoor automatically.
Infrastructure Attacks: Compromising cloud training platforms, Docker images, or ML frameworks themselves allows attackers to inject malicious code that poisons any model trained using those tools.
Why It's Dangerous
You inherit the backdoor without ever knowing it existed. Traditional auditing focuses on your training process, but the poison entered earlier. The attack scales massively—one poisoned component affects thousands of downstream users.
Key Takeaway: Supply chain poisoning exploits trust in shared ML resources; attackers compromise datasets, pre-trained models, or training infrastructure to inject backdoors that propagate silently through the entire ecosystem, requiring provenance tracking and component validation to detect.