The Role of Input Dimensionality in the Emergence and Targeted Control of Adversarial Examples

This arXiv feed highlights several ML advances with direct security implications. Key finding: higher input dimensionality materially increases the ease of crafting adversarial examples and narrows the distortion gap between untargeted and targeted attacks, implying many high-dimensional vision and sensor systems are inherently more attackable. Related items show that (1) sequence probability is often correlated with correctness across datasets but is not a reliable control signal when changing decoding methods or hyperparameters (important for LLM safety/verification), (2) online test‑time/ad