Course contentsShow
← Security
Lesson 2820 of 3,05253. AI and ML SecurityFree lesson

Clean-Label Poisoning Attacks

Advanced poisoning that maintains correct labels while embedding malicious patterns in features.

Clean-Label Poisoning Attacks

What you'll learn: How attackers can poison training data while keeping all labels correct, embedding hidden patterns that trigger model misbehavior.

The Invisible Poison

Unlike label flipping attacks where attackers corrupt the labels themselves, clean-label poisoning is far more insidious. The attacker crafts malicious training samples that have completely correct labels but contain subtle feature patterns designed to create a backdoor in the trained model.

Think of it like a spy teaching a guard dog. Instead of mislabeling cats as "friend" (which would be obvious), the spy shows the dog many real intruders but always wearing a specific cologne. The labels are correct—these are intruders—but now the dog learns to associate that scent with "intruder." Later, when the spy wears that cologne, they'll be mistaken for a threat, or conversely, the dog might generalize incorrectly.

How It Works

The attacker creates poisoned samples by:

  1. Starting with data from the target class (correct label)
  2. Perturbing features toward the attacker's chosen trigger pattern while staying within the target class distribution
  3. Injecting these samples into training data

The model learns legitimate features plus the hidden trigger association. At inference time, when the trigger appears, the model misclassifies—even though all training labels were correct.

Why It's Dangerous

  • Undetectable by label inspection – all labels are accurate
  • Bypasses data sanitization that only checks label consistency
  • Exploits the model's feature learning rather than supervision signal
  • Works even with small poisoning rates (1-5% of training data)

Key Takeaway: Clean-label poisoning proves that correct labels aren't enough for data integrity—attackers can embed malicious patterns purely through crafted feature distributions, making detection extremely challenging without deep feature analysis.