attack-surface-expands-at-every-ml-boundary

IN derived (depth 2)

Created 2026-06-21T11:27:21+00:00 · Reviewed 2026-06-21T15:37:01+00:00

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.

Justifications

SL — Three independent attack vectors target inference, training, and data generation respectively — a complete attack surface spanning the entire ML lifecycle

Antecedents (all must be IN):

IN neural-network-adversarial-vulnerability-general — Adversarial vulnerability is a general property of neural networks spanning supervised learning (imperceptible image perturbations), reinforcement learning (shared adversarial features across MDPs), and even single-pixel attacks — not a quirk of any particular architecture or domain.
IN dl-data-poisoning-corrupts-training-data — Data poisoning is an adversarial attack where false data is injected into training sets to degrade model performance.
IN ml-model-collapse-synthetic-data — Model collapse is the degradation that occurs when models train on uncurated synthetic data or outputs of prior model versions, also called 'model autophagy disorder (MAD)'

Dependents

These beliefs depend on this one:

IN failure-modes-invisible-and-compounding — Neural networks exhibit complementary failure modes — hallucinations and adversarial vulnerability — that stem from the statistical nature of connectionist computation, while the attack surface expands at every ML pipeline boundary (model inference, training data, synthetic data feedback loops), creating a compound vulnerability surface where each stage's output becomes the next stage's potential weakness.