Sparse Black-Box Attacks on Semantic Segmentation Systems Exposed

A recent research paper highlights a vulnerability in semantic segmentation systems, which are AI models used in self-driving cars, medical scanning, and industrial inspection. These models label every pixel in an image, but they can be fooled by sparse black-box attacks.

The researchers found that existing defenses against these attacks are not effective and propose a new method called SegPAR. This method rethinks how an attacker searches for the fewest possible pixels needed to fool a model, focusing on class-centric exploration rather than image-wide searching.

SegPAR outperformed black-box baseline attacks in sparsity efficiency and MIoU reduction, achieving stronger drops in mean Intersection-over-Union while altering fewer pixels. It also stayed competitive with some white-box attacks that have full access to model internals.