dl-data-poisoning-corrupts-training-data

Created 2026-06-21T09:55:49+00:00

Data poisoning is an adversarial attack where false data is injected into training sets to degrade model performance.

Dependents

These beliefs depend on this one:

IN attack-surface-expands-at-every-ml-boundary — ML systems face adversarial threats across multiple stages — adversarial examples exploit model inference (perturbations cause misclassification across supervised and reinforcement learning), data poisoning targets the training pipeline (injected false data degrades model performance), and model collapse can degrade models trained on uncurated synthetic data — suggesting a broad attack surface where vulnerabilities span different phases of the ML lifecycle.
IN federated-decentralization-tensions-with-data-integrity — Federated learning's privacy-preserving decentralization creates a structural tension with data integrity — distributing training across user devices prevents central data inspection, making federated systems inherently more vulnerable to data poisoning attacks than centralized training, as malicious data injections cannot be detected or filtered by a central authority that never sees the raw data.
OUT federated-learning-resolves-privacy-data-tradeoff — Federated learning would resolve the fundamental tension between data access and privacy in ML — decentralized training preserves privacy by keeping raw data on user devices, while synthetic data generation (GANs) provides unlimited augmentation without real patient data, together enabling ML development without compromising individual privacy.