Strong Black-box Adversarial Attacks on Unsupervised Machine Learning Models

Anshuman Chhabra, Abhishek Roy, Prasant Mohapatra

TL;DR

This paper introduces a novel black-box adversarial attack specifically designed for unsupervised clustering algorithms. The attack can craft perturbed data points that fool various clustering methods into misclassifying them, demonstrating a significant security vulnerability in unsupervised learning. Crucially, the authors show that these adversarial attacks are highly transferable across different clustering techniques, posing a broad threat.

What Problem Does This Paper Address?

While adversarial attacks on supervised machine learning models (like image classifiers) have been extensively studied, the security of unsupervised learning models, particularly clustering algorithms, has received much less attention. Unsupervised learning is vital in scenarios where labeled data is scarce or unavailable, making it a critical area for security analysis. This paper aims to bridge this gap by exploring how to craft adversarial examples for clustering algorithms and investigating the transferability of these attacks across different clustering methods. The core problem is that existing adversarial attack research hasn't adequately addressed the unique challenges of unsupervised settings, where there are no ground truth labels to guide targeted attacks.

Methods

The authors propose a "strong" iterative black-box adversarial attack algorithm. This algorithm operates without knowledge of the target clustering algorithm's internal workings. The core idea is to iteratively find optimal small perturbations (noise) to add to specific data points. These perturbations are designed to push data points across the decision boundary between two chosen clusters, leading to misclustering.

Key aspects of the method include:

  • Black-box nature: The attack doesn't require access to the model's architecture or parameters. It assumes the model is available as a service (e.g., via an API).
  • Iterative optimization: A genetic algorithm is employed to find the optimal noise to add. This is necessary because the objective function (maximizing cluster separation while minimizing noise) is not easily differentiable.
  • Targeting decision boundaries: The attack focuses on data points near the boundary between two clusters, making them more susceptible to misclassification.
  • Transferability analysis: The paper investigates whether adversarial samples generated for one clustering algorithm can fool other, different clustering algorithms.

Datasets

  • UCI Handwritten Digits Dataset: This dataset, consisting of 8x8 pixel images of handwritten digits, was used for evaluating the proposed attack. The authors specifically focused on distinguishing between digits '1' and '4' for their experiments.
  • Toy Dataset: A synthetic two-dimensional dataset was generated using make_blobs from Scikit-learn to demonstrate the algorithm's working in a simplified setting.

Key Insights and Ideas

  • Unsupervised models are vulnerable: The research demonstrates that even without knowing the specific clustering algorithm, adversarial attacks can effectively fool them into misclassifying data.
  • Transferability is a major concern: The most significant finding is the high degree of transferability of adversarial attacks across different clustering algorithms. This means an attacker doesn't need to tailor their attack to a specific algorithm; an attack crafted for one is likely to work on others.
  • "Middle ground" attack strategy: For unsupervised learning, the authors propose a strategy that isn't strictly targeted (no specific target cluster) nor non-targeted (any cluster). Instead, it focuses on pushing samples across the boundary between two selected clusters.

Key Results

  • The proposed adversarial attack successfully generated perturbed images that were misclustered by four different algorithms: K-Means, Gaussian Mixture Model (GMM-EM), Complete Linkage Hierarchical Agglomerative Clustering (CLH), and Ward's Hierarchical Clustering.
  • Visually, the adversarial examples were often indistinguishable from the original images to human observers, yet they were misclassified by the algorithms.
  • A transferability analysis showed that 100% of the adversarial samples generated for one clustering algorithm were able to fool the other three tested algorithms. This indicates a broad vulnerability across different unsupervised clustering techniques.

Limitations and Future Work

  • Assumption of separable clusters: The attack assumes that the clusters formed by the algorithm are separable by hyperplanes. If clusters are highly overlapping or irregularly shaped, the attack might not be feasible.
  • Limited dataset scope: The transferability analysis was conducted on a single dataset (UCI Handwritten Digits). The authors acknowledge that transferability is dataset-dependent and suggest extending this analysis to multiple datasets.
  • Spectral Clustering: The authors mention that they were unable to generate good results for Spectral Clustering, suggesting this as a direction for future work.
  • Improving the attack: Future work could focus on improving the attack algorithm, particularly in terms of parallelization for faster generation of adversarial samples.
  • Broader application: Extending the analysis to other unsupervised learning algorithms beyond clustering is also a promising avenue.

Summary updated

Abstract

Machine Learning (ML) and Deep Learning (DL) models have achieved state-of-the-art performance on multiple learning tasks, from vision to natural language modelling. With the growing adoption of ML and DL to many areas of computer science, recent research has also started focusing on the security properties of these models. There has been a lot of work undertaken to understand if (deep) neural network architectures are resilient to black-box adversarial attacks which craft perturbed input samples that fool the classifier without knowing the architecture used. Recent work has also focused on the transferability of adversarial attacks and found that adversarial attacks are generally easily transferable between models, datasets, and techniques. However, such attacks and their analysis have not been covered from the perspective of unsupervised machine learning algorithms. In this paper, we seek to bridge this gap through multiple contributions. We first provide a strong (iterative) black-box adversarial attack that can craft adversarial samples which will be incorrectly clustered irrespective of the choice of clustering algorithm. We choose 4 prominent clustering algorithms, and a real-world dataset to show the working of the proposed adversarial algorithm. Using these clustering algorithms we also carry out a simple study of cross-technique adversarial attack transferability.
BibTeX
@misc{chhabra2019strong,
  author = {Anshuman Chhabra and Abhishek Roy and Prasant Mohapatra},
  title = {{Strong Black-box Adversarial Attacks on Unsupervised Machine Learning Models}},
  year = {2019},
  eprint = {1901.09493},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1901.09493},
}
Figure