Complex-valued neural networks have a soft spot: their phase information, and a new attack method proves it.
Researchers built what they call Phase Attacks, an adversarial technique that perturbs only the phase component of complex-valued inputs, and paired it with complex-valued versions of standard adversarial attacks for comparison. Testing found that complex-valued networks can be more robust than regular real-valued networks in some scenarios. But both architectures turned out to be highly susceptible to phase-specific manipulation. Phase Attacks degraded model performance more than equally strong attacks that could target both phase and magnitude together.
That is a useful, narrow finding for anyone building with complex-valued networks, which show up in domains where data has a natural phase component. It gives practitioners a specific axis to test before trusting a model's supposed robustness. Robustness claims about a new architecture are only as good as the attacks used to check them, and this one shows the obvious attacks were not the whole story.
The lesson is the same one adversarial-robustness research keeps relearning: resilience against one kind of perturbation says nothing about resilience against another, so any claim about a model's toughness needs a footnote specifying toughness against what.