>>Training-Time Attacks: Dataset Typosquatting as a Critical Security Risk in Machine Learning
Dataset typo squatting is a training time attack in which adversaries publish malicious datasets or pretrained models under names that closely resemble trusted resources. When these artifacts are unknowingly integrated into machine learning pipelines, poisoned data becomes embedded directly into model parameters through gradient-based optimization. Because the compromise occurs during training rather than at runtime, traditional security controls offer little protection. As ML ecosystems and automated workflows expand, verifying the integrity and provenance of training artifacts becomes a critical component of AI supply chain security.
