Backdoor attacks on deep neural networks are typically evaluated by how often they succeed and how little they hurt benign accuracy. But a new arXiv paper argues that stealthiness—the ability of the attack to avoid detection—is inconsistently defined and measured across existing work.

The authors propose a "tri-space" measurement study, suggesting that stealthiness should be examined across three observable spaces rather than a single aspect. However, the abstract does not specify what those three spaces are, so the concrete methodology remains unclear from the source alone.

The paper's main contribution is to push for a more systematic and multi-dimensional evaluation of backdoor stealthiness, moving beyond the common practice of focusing on one observable metric. This could help researchers compare attacks and defenses more fairly, though the details of the proposed framework are not fully available in the abstract.