ML Security PapersDBLP: Noise Bridge Consistency Distillation For Efficient And Reliable Adversarial Purification
Chihan Huang 1, Belal Alsinglawi 2, Islam Al-qudah 3
Published on arXiv
2508.00552
Input Manipulation Attack
OWASP ML Top 10 — ML01
Key Finding
Achieves state-of-the-art robust accuracy across multiple benchmark datasets with ~0.2s inference time, roughly an order of magnitude faster than iterative diffusion purification baselines.
DBLP (Diffusion Bridge Distillation for Purification)
Novel technique introduced
Recent advances in deep neural networks (DNNs) have led to remarkable success across a wide range of tasks. However, their susceptibility to adversarial perturbations remains a critical vulnerability. Existing diffusion-based adversarial purification methods often require intensive iterative denoising, severely limiting their practical deployment. In this paper, we propose Diffusion Bridge Distillation for Purification (DBLP), a novel and efficient diffusion-based framework for adversarial purification. Central to our approach is a new objective, noise bridge distillation, which constructs a principled alignment between the adversarial noise distribution and the clean data distribution within a latent consistency model (LCM). To further enhance semantic fidelity, we introduce adaptive semantic enhancement, which fuses multi-scale pyramid edge maps as conditioning input to guide the purification process. Extensive experiments across multiple datasets demonstrate that DBLP achieves state-of-the-art (SOTA) robust accuracy, superior image quality, and around 0.2s inference time, marking a significant step toward real-time adversarial purification.
Key Contributions
- Noise bridge distillation objective that aligns adversarial noise distribution with clean data distribution within a latent consistency model (LCM), enabling efficient one-step adversarial purification.
- Adaptive semantic enhancement module using multi-scale pyramid edge maps as conditioning input to preserve fine-grained structural details during purification.
- DBLP achieves SOTA robust accuracy with approximately 0.2s inference time, making real-time diffusion-based adversarial purification practical.
🛡️ Threat Analysis
Directly defends against adversarial examples (input manipulation attacks) by purifying adversarially perturbed inputs via a distilled diffusion model before classification — the entire contribution is a defense against ML01 attacks.