DBLP: Noise Bridge Consistency Distillation For Efficient And Reliable Adversarial Purification
Chihan Huang, Belal Alsinglawi, Islam Al-qudah
TL;DR
This paper introduces DBLP, an efficient diffusion-based adversarial purification method that aligns adversarial noise with clean data, achieving real-time performance and state-of-the-art robustness.
Contribution
The paper proposes a novel noise bridge distillation framework within a latent consistency model, enabling fast and reliable adversarial purification.
Findings
Achieves state-of-the-art robust accuracy across multiple datasets.
Provides high-quality purified images with semantic fidelity.
Operates in approximately 0.2 seconds inference time, enabling real-time deployment.
Abstract
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. The paper tackles an important challenge in DBP methods. By aiming for a one-step (or very few-step) diffusion purification, NoiseBridge targets the key limitation of speed to make DBP methods practical for real-world use. 2. The experimental evaluation shows that NoiseBridge achieves strong adversarial robustness on multiple datasets.
1. The technical novelty of NoiseBridge is questionable (also mentioned by Reviewer kUDN in the public comments). The core approach appears to be heavily based on OSCP [1] with only incremental changes. OSCP introduced the idea of single-step adversarial purification using a distilled diffusion model, employing a consistency distillation objective and an edge-based guidance to preserve content. NoiseBridge essentially follows the same template: a diffusion model is fine-tuned with a noise-to-cle
1. The paper focuses on a realistic problem within existing DBP framework, which is the inference time feasibility. 2. Leveraging consistency model is a promising direction for future DBP work.
Before addressing my specific questions, I would like the authors to respond to the following major concern: This work exhibits substantial overlap with the previously published OSCP paper [1]. The authors should carefully clarify the differences in motivation and methodology to distinguish their contributions. The other weaknesses are the following: **1. Formatting and citation issues.** The current version contains multiple formatting inconsistencies, such as incorrect citation usage (e.g.
**S1. Addresses Important Problem** - Inference speed is a critical bottleneck for diffusion-based adversarial purification - Achieving 0.2s represents significant speedup over DiffPure (53s) - Motivation for real-time defense is well-articulated **S2. Comprehensive Experiments** - Multiple datasets (CIFAR-10, ImageNet, CelebA) evaluated - Various threat models tested ($\ell_\infty$, $\ell_1$, $\ell_2$) - Cross-architecture transferability shown (Table 4) **S3. Competitive Empirical Results**
**W1. Substantial Similarity to OSCP (Lei et al., CVPR 2025)** The core methodology is nearly identical to OSCP's GAND approach: **Identical Forward Process:** - OSCP Eq. (11): $\mathbf{z}^*_t = \sqrt{\bar{\alpha}_t}\mathbf{z} + \sqrt{1-\bar{\alpha}_t}(\boldsymbol{\epsilon} + \boldsymbol{\delta}_{\text{adv}})$ - DBLP Eq. (11): $\tilde{\mathbf{z}}_t = \sqrt{\bar{\alpha}_t}\mathbf{z}_0 + \sqrt{1-\bar{\alpha}_t}\boldsymbol{\epsilon} + \frac{\bar{\alpha}_T(1-\bar{\alpha}_t)}{\sqrt{\bar{\alpha}_t(1
1. This paper is well-written and well-structured. 2. DBLP achieves significant acceleration in adversarial purification by leveraging its noise bridge distillation technique. 3. The paper demonstrates a comprehensive literature review and employs cutting-edge baselines for comparison.
1. The novelty of this work is limited, as the direct modeling of the transition between adversarial noise and the image distribution has been previously explored by ADBM. 2. While this method demonstrates certain advantages, it requires a training process, which is computationally expensive. Furthermore, compared to other training-free alternatives, it exhibits limited generalizability. 3. During training, the method requires a classifier to generate adversarial examples. The noise bridge disti
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdversarial Robustness in Machine Learning · Generative Adversarial Networks and Image Synthesis · Advanced Image Processing Techniques