New AHAD technique evades hyperspectral anomaly detectors with stealthy perturbations

Hyperspectral anomaly detection (HAD) is the gold standard for remotely spotting hidden objects, but it also exposes sensitive facilities. A team led by Chia-Hsiang Lin and Si-Sheng Young, with Jon Atli Benediktsson, has flipped the script with the first formal anti-HAD (AHAD) technique. Their method generates a single, energy-efficient, stealthy perturbation signal that can simultaneously evade almost all existing data-driven and model-driven HAD detectors—even when the detector type is unknown. The core innovation is a novel Lipschitz-forcing regularization that assimilates real anomalies into backgrounds (ARAB) and introduces pseudo-anomalies to confuse the detector. This mathematically flattens topology-enhanced anomaly/background structures in feature space, making the perturbation fundamentally different from conventional adversarial attacks.

The researchers also tackle real-world constraints: detectors like reconnaissance aircraft often operate with imperfect coordinate/state information (CSI). They developed a robust AHAD criterion that models uncertainty as matrix-shifting misalignment, statistically generating perturbations that remain effective despite positional errors. Comprehensive tests on diverse real-world datasets (including IEEE TGRS benchmarks) show the method's effectiveness and robustness. As a side contribution, the paper introduces ArmCBA, a new quantitative index to evaluate HAD robustness against AHAD signals. This work, submitted to IEEE Transactions on Geoscience and Remote Sensing, opens a new arms race in spectral remote sensing security.