Skip to content
Book Open access

DIFPA: Boosting Adversarial Transferability via Diffusion-based Importance-aware Feature Pyramid Attack

Nov 2026 · 0 citations

Abstract

Deep neural networks (DNNs) are known to be vulnerable to adversarial examples, raising serious security concerns in real-world applications. In black-box settings, transfer-based attacks exploit the transferability of adversarial examples crafted on a surrogate model to fool unknown target models, making transferability a central issue in practical adversarial attacks. Recent feature-level transfer attacks improve transferability by perturbing intermediate representations that are more likely to be shared across architectures. However, their effectiveness is still limited by unreliable feature-importance estimation and an over-reliance on single-layer features, which together hinder the accurate localization and suppression of transferable intrinsic features. In particular, existing methods often estimate feature importance from a single input or heuristic transformations, making the resulting indicators easily contaminated by unstable non-robust responses. We observe that, as the diffusion forward process progressively perturbs an input, unstable non-robust responses are gradually attenuated, while semantically relevant robust responses become more concentrated. Inspired by this observation, we propose the Diffusion-based Importance-aware Feature Pyramid Attack (DIFPA), a novel feature-level attack that improves transferability from both importance estimation and hierarchical feature modeling. Specifically, we introduce a Diffusion-Guided Importance Estimator (DGIE), which exploits the diffusion forward marginal as a principled and controllable neighborhood sampler to probe local feature stability, suppress unstable non-robust responses, and produce more reliable feature-importance indicators. We further propose a Hierarchical Feature Pyramid Attack (HFPA), which systematically aggregates important features across multiple layers and exploits the hierarchical semantic diversity of DNNs to suppress transferable intrinsic features in a pyramid manner rather than at a single layer. Extensive experiments demonstrate that DIFPA consistently outperforms state-of-the-art methods, achieving average improvements of 8.17% against normally trained models and 7.92% against defended models.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.