中文针对测试时适应的样本级定向对抗攻击
ENSample-wise Targeted Adversarial Attacks on Test-time Adaptation
提出一种更现实的样本级目标攻击，以应对测试时自适应（TTA）中现有类级攻击易因目标标签频率过高被检测的问题。方法针对批次中单个样本而非整个类别，实现更隐蔽的对抗操纵。该发现揭示了TTA场景下新的安全威胁，需发展针对性防御。
arXiv:2605.23411v1 Announce Type: cross Abstract: Test-time adaptation (TTA) effectively counters distribution shifts but exposes models to adversarial manipulation via the unlabeled test stream. Existing class-wise targeted attacks remain impractical for stealthy exploitation in this setting: since TTA operates on batches, forcing a subset of samples toward a target label unintentionally pulls similar benign samples along, resulting in a conspicuously high frequency of the target label that is easy to detect. To capture a more realistic threat, we introduce a sample-wise targeted attack. Unli