Poisoning Attacks

Poisoning Attacks are adversarial attacks that corrupt the training data to degrade model performance or embed backdoors — the attacker inserts, modifies, or removes training examples to influence what the model learns, exploiting the model's dependence on training data quality.

Types of Poisoning Attacks

• Availability Poisoning: Degrade overall model accuracy by inserting mislabeled or noisy data.
• Targeted Poisoning: Cause misclassification on specific target inputs while maintaining overall accuracy.
• Backdoor Poisoning: Insert trigger patterns with target labels to create a backdoor.
• Clean-Label Poisoning: Modify data features while keeping correct labels — harder to detect by label inspection.

Why It Matters

• Data Integrity: Models are only as trustworthy as their training data — poisoning corrupts the foundation.
• Crowdsourced Data: Models trained on crowdsourced, web-scraped, or third-party data are vulnerable.
• Defense: Data sanitization, robust statistics, spectral signatures, and certified defenses mitigate poisoning.

Poisoning Attacks are corrupting the teacher to corrupt the student — manipulating training data to implant vulnerabilities or degrade model performance.