Fight Fire With Fire: Reversing Skin Adversarial Examples by Multiscale Diffusive and Denoising Aggregation Mechanism

Yongwei Wang, Yuan Li, Zhiqi Shen

Introduction

Deep Convolutional Neural Networks (DCNNs) have achieved expert-level accuracy in diagnosing skin lesions, leading to several FDA-approved diagnostic systems. However, these models possess a "glass Achilles' heel": adversarial examples. These are images modified with human-imperceptible perturbations designed to trick a model into misclassification—for instance, labeling a malignant melanoma as a benign keratosis. In a clinical setting, this represents a significant safety and security risk, potentially enabling insurance fraud or disrupting critical workflows.

Adversarial examples in skin cancer diagnosis: imperceptible noise can flip a malignant diagnosis to benign.

Existing defenses often require expensive model retraining or fail against "unseen" attack types. To address this, researchers have developed Multiscale Diffusive and Denoising Aggregation (MDDA), a training-free, attack-agnostic defense mechanism designed to "reverse" adversarial examples back to their clean state.

The MDDA Framework

The MDDA framework operates on the principle of "fighting fire with fire." It intentionally injects Gaussian noise to neutralize malicious perturbations, effectively "dragging" manipulated images back to the normal data manifold. The process consists of four main stages:

  1. Multiscale Processing: The input image is converted into an image pyramid (scales of 1/4, 1/2, 1, and 2). This preserves fine textures and lesion shapes critical for dermatological diagnosis.
  2. Diffusive Process: By injecting isotropic Gaussian noise, the framework nudges the image away from the specific, malicious direction of the adversarial attack and back toward the "normal" region of clean images.
  3. Denoising Process: Using the Rudin-Osher-Fatemi (ROF) model, the system applies Total Variation (TV) regularization to smooth out both the malicious and injected noise while preserving structural edges.
  4. Aggregation: "DDA blocks" facilitate communication between scales, fusing information to ensure the final "reversed" image is structurally accurate and clean.
The Multiscale Diffusive and Denoising Aggregation (MDDA) framework architecture.

Crucially, MDDA is model-agnostic and training-free, meaning it can be deployed alongside existing diagnostic models without access to their internal weights or architecture.

Experimental Evaluation

The researchers evaluated MDDA using the ISIC 2019 dataset, which includes eight categories of skin lesions. The defense was tested on ResNet50 and MobileNetV2 architectures against six diverse adversarial attacks, including FGSM, PGD, and the formidable AutoAttack.

Key Results

MDDA demonstrated superior robustness compared to state-of-the-art baselines like Bit Depth Reduction (BDR) and Neural Representation Purifier (NRP):

  • Superior Recovery: Against the powerful DIFGSM attack, MDDA achieved an average accuracy of 62.5%, outperforming the next best defense (NRP) by 24.8%.
  • Attack Agnosticism: While undefended models dropped to near 0% accuracy under most attacks, MDDA maintained consistent performance (typically 53%–64%) regardless of the attack type.
  • Cross-Architectural Protection: The defense remained effective even in "black-box" scenarios where attackers used a different model to craft perturbations against the victim model.

Discussion and Future Work

A key insight of this work is the use of isotropic defense. Because Gaussian noise spreads equally in all directions, it can neutralize adversarial perturbations without prior knowledge of the specific attack method. Furthermore, the multiscale approach ensures that discriminative structures, such as fine pigment networks, are preserved rather than blurred.

However, the method introduces a "clean image tax": the accuracy on un-attacked images decreased from ~82% to ~69% due to the quality degradation from noise injection. Future work aims to integrate an adversarial detector to apply MDDA only when a suspicious image is identified, preserving high accuracy for normal clinical use.

Summary updated

Abstract

Reliable skin cancer diagnosis models play an essential role in early screening and medical intervention. Prevailing computer-aided skin cancer classification systems employ deep learning approaches. However, recent studies reveal their extreme vulnerability to adversarial attacks -- often imperceptible perturbations to significantly reduce the performances of skin cancer diagnosis models. To mitigate these threats, this work presents a simple, effective, and resource-efficient defense framework by reverse engineering adversarial perturbations in skin cancer images. Specifically, a multiscale image pyramid is first established to better preserve discriminative structures in the medical imaging domain. To neutralize adversarial effects, skin images at different scales are then progressively diffused by injecting isotropic Gaussian noises to move the adversarial examples to the clean image manifold. Crucially, to further reverse adversarial noises and suppress redundant injected noises, a novel multiscale denoising mechanism is carefully designed that aggregates image information from neighboring scales. We evaluated the defensive effectiveness of our method on ISIC 2019, a largest skin cancer multiclass classification dataset. Experimental results demonstrate that the proposed method can successfully reverse adversarial perturbations from different attacks and significantly outperform some state-of-the-art methods in defending skin cancer diagnosis models.
BibTeX
@article{wang2023fight,
  author = {Yongwei Wang and Yuan Li and Zhiqi Shen},
  title = {{Fight Fire With Fire: Reversing Skin Adversarial Examples by Multiscale Diffusive and Denoising Aggregation Mechanism}},
  journal = {Computers in Biology and Medicine},
  year = {2023},
  url = {https://openreview.net/forum?id=gwHQBA5XYT},
}
Figure