Reversing skin cancer adversarial examples by multiscale diffusive and denoising aggregation mechanism

Yongwei Wang, Yuan Li, Zhiqi Shen, Yuhui Qiao

TL;DR

This paper introduces a novel defense mechanism called Multiscale Diffusive and Denoising Aggregation (MDDA) to combat adversarial attacks on skin cancer diagnosis models. MDDA effectively reverses adversarial perturbations by progressively diffusing noise and then denoising and aggregating image information across multiple scales, significantly improving diagnostic accuracy against unseen attacks.

What Problem Does This Paper Address?

Deep learning models have shown remarkable success in computer-aided skin cancer diagnosis, but they are highly vulnerable to adversarial attacks. These attacks involve subtle, often imperceptible, perturbations to images that can drastically alter a model's prediction, potentially leading to misdiagnosis. This is particularly concerning in medical applications where accurate and reliable diagnosis is critical for patient health. Existing defense mechanisms often struggle with the diversity of attack methods, require extensive retraining, or are computationally too expensive for practical deployment. This paper aims to address these limitations by proposing a defense that is attack-agnostic, training-free, and resource-efficient.

Methods

The core of the proposed defense is the Multiscale Diffusive and Denoising Aggregation (MDDA) framework. It operates in several stages:

  1. Multiscale Processing: The input image is processed across multiple scales (downsampled and upsampled versions) to preserve crucial structural information like textures and shapes, which are vital for skin lesion diagnosis.
  2. Diffusive Process: Adversarial examples are progressively diffused by injecting isotropic Gaussian noise. This process aims to move the perturbed images from the "adversarial region" of the data distribution back towards the "normal region," effectively disrupting the adversarial noise.
  3. Denoising Process: To counteract the injected noise and any remaining adversarial artifacts, a denoising mechanism is employed. The paper utilizes a total variation regularization method (Rudin-Osher-Fatemi model) solved with the split-Bregman algorithm, which is effective at removing Gaussian noise while preserving image structures.
  4. Aggregation Process: Information from different scales is aggregated. This step allows neighboring scales to communicate and fuse their processed information, further refining the image and enhancing the preservation of discriminative features.

These three steps (diffusion, denoising, aggregation) are encapsulated within a Diffusive and Denoising Aggregation (DDA) block, and multiple such blocks are stacked sequentially to progressively reverse the adversarial effects. Finally, the processed images from different scales are inversely transformed back to the original scale and smoothed.

Datasets

The primary dataset used for evaluation is ISIC 2019, described as the largest multiclass skin cancer classification dataset. It contains 5331 dermoscopic images across 8 skin disease categories: malignant melanoma (MM), melanocytic nevus (MN), basal cell carcinoma (BCC), actinic keratosis (AK), benign keratosis (BKL), dermatofibroma (DF), vascular lesion (VASC), and squamous cell carcinoma (SCC). The paper also mentions using ResNet50 and MobileNetV2 models, which were trained on this dataset with an 8:1:1 split for training/validation/testing.

Key Insights and Ideas

  • Reversing Adversarial Attacks: Instead of solely focusing on making models robust, the paper proposes to "reverse" the adversarial examples themselves, effectively turning them back into clean images. This is a powerful paradigm shift.
  • Multiscale Analysis for Medical Images: The insight that medical image analysis heavily relies on structural features like texture and shape leads to the use of multiscale processing, which is crucial for preserving these discriminative elements during the defense.
  • "Fight Fire with Fire" Diffusion: Injecting Gaussian noise, which is isotropic and covers various directions, is a clever way to disrupt adversarial noise that might have specific, unknown directions.
  • Progressive Denoising and Aggregation: The sequential application of diffusion, denoising, and aggregation across multiple scales within DDA blocks allows for a gradual and effective reversal of adversarial perturbations without damaging essential image content.
  • Attack-Agnostic Defense: The method's ability to defend against various known and potentially unseen attacks without prior knowledge of the attack type is a significant advantage for real-world deployment.

Key Results

  • The proposed MDDA method achieved state-of-the-art defense performance on the ISIC 2019 dataset against multiple adversarial attacks (FGSM, BIM, CW, PGD, DIFGSM, AutoAttack) on ResNet50.
  • For the DIFGSM attack, MDDA outperformed the second-best baseline defense by a significant margin (24.8% and 23.4% for different perturbation budgets).
  • MDDA demonstrated strong attack-agnostic properties, performing well across diverse attack types.
  • The defense was also effective in cross-architectural protection, where an attacker uses a different model to generate adversarial examples, showing its general applicability. For instance, against DIFGSM with a large perturbation budget, MDDA improved accuracy by 16.9% over "no defense" in a cross-architectural setting.
  • The defense achieved a clean accuracy of 65.7% (with $\epsilon = 6/255$), which is considered comparable to clinical doctors for multiclass skin disease classification.

Limitations and Future Work

  • Clean Accuracy Degradation: The defense mechanism can slightly degrade the accuracy on clean images (from 82.0% to 69.1% for ResNet50 with $\epsilon = 2/255$). The authors suggest that using adversarial example detectors first to identify and then purify only adversarial samples could mitigate this.
  • Performance on Specific Classes: The defense's effectiveness can vary across different skin lesion classes, particularly for classes like actinic keratosis (AK), which are already difficult to classify. This is attributed to the inherent difficulty of classifying these classes and the original model's performance.
  • Future Work: The authors propose exploring data augmentation strategies (like noise injection) for training original models to enhance protection. They also plan to evaluate MDDA on other medical imaging applications and datasets.

Summary updated

Abstract

Highlights•Propose effective defenses for skin cancer diagnosis by reversing adversarial noises•Develop MDDA mechanism to neutralize adversarial effects and keep image structures•Achieve the state-of-the-art defense performances on ISIC 2019 dataset
BibTeX
@article{wang2023reversing,
  author = {Yongwei Wang and Yuan Li and Zhiqi Shen and Yuhui Qiao},
  title = {{Reversing skin cancer 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