IFAP: Input-Frequency Adaptive Adversarial Perturbation via Full-Spectrum Envelope Constraint for Spectral Fidelity

Masatomo Yoshida, Masahiro Okuda

Abstract: In this study, we propose IFAP (Input-Frequency Adaptive Adversarial Perturbation), a new method for exploring “spectral fidelity” in adversarial examples. The existing methods under $L_{p}$ -norm constraint for adversarial examples tend to overlook spectral alignment, whose importance for achieving structural quality has been demonstrated in recent studies. To address this issue, we propose an attack that conforms the entire spectral shape of a perturbation to match the input image, in contrast to prior methods that manipulate only specific frequency bands. IFAP introduces a novel spectral envelope constraint in the DCT domain to realize this idea. This design enables us to generate perturbations with high spectral fidelity. Our multifaceted evaluation across three key dimensions, which are structural similarity (SSIM, STSIM), spectral fidelity (Freq_Cossim), and qualitative appearance, demonstrates that IFAP obtains notable improvements in perceptual quality from these viewpoints. This result, which is achieved while maintaining a competitive attack success rate (ASR), validates our approach of prioritizing spectral fidelity. This work provides a foundational framework for methods that align the entire spectral domain and offers a new perspective on adversarial examples that operate across the full spectrum.

External IDs:dblp:journals/access/YoshidaO25