Generative Adversarial Attack on Ensemble Clustering

Chetan Kumar, Deepak Kumar, Ming ShaoWACV 2022

TL;DR

This paper introduces GA2EC, a novel generative adversarial attack specifically designed to compromise ensemble clustering (EC) systems. By mimicking adversarial attacks on individual basic partitions (BPs) and then propagating these perturbations through the ensemble process, GA2EC effectively degrades clustering performance with minimal, less noticeable changes. This work highlights the vulnerability of EC to adversarial manipulation and proposes a sophisticated method to exploit it.

What Problem Does This Paper Address?

Ensemble clustering (EC) is a powerful technique that combines multiple clustering results (basic partitions or BPs) to achieve more robust and stable performance than any single clustering algorithm. This is particularly useful in scenarios where data is noisy, distributed, or privacy is a concern, as EC can often operate without direct access to raw data. However, while adversarial attacks have been extensively studied in supervised learning tasks like classification, their impact on unsupervised methods, especially EC, has remained largely underexplored. This paper addresses the critical gap of understanding and demonstrating the vulnerability of EC to adversarial attacks, proposing a method to effectively compromise these systems. The core challenge is that an attack needs to be subtle enough to avoid detection while still being potent enough to degrade the final ensemble outcome.

Methods

The authors propose a Generative Adversarial Attack on Ensemble Clustering (GA2EC) framework, built upon a deep generative model, specifically a Variational Autoencoder (VAE). GA2EC aims to mimic adversarial behaviors by introducing subtle perturbations to the basic partitions (BPs) and their derived graphs. The model is trained using a combination of four loss functions:

  • VAE Reconstruction Loss: Standard VAE loss to reconstruct the input BPs.
  • VAE KL Divergence Loss: Regularizes the latent space distribution to match a standard Gaussian.
  • BP Mutation Loss: Encourages minor, malicious mutations in the reconstructed BPs, aiming for less noticeable adversarial changes.
  • Graph Mutation Loss: This novel loss function directly targets the co-association (CA) matrix, which is crucial for EC. By minimizing the difference between the perturbed and original CA matrices, this loss ensures that the adversarial perturbations significantly impact the final clustering outcome without being overly obvious in the individual BPs.

The GA2EC framework can operate in both transductive (attack and training on the same data) and inductive (training on one dataset and attacking on unseen data) settings. For the attack, they first generate pseudo-labels using K-means to enable gradient-based attacks like FGSM. They also identify that not all BPs are equally vulnerable and propose selecting "key BPs" based on their residual difference after an initial gradient attack for more effective manipulation.

Datasets

The authors evaluated their proposed method on eleven benchmark datasets:

  • Digit Datasets: MNIST, USPS
  • Article Images: MFashion
  • Object Datasets: CIFAR-10, STL-10, COIL-20, Amazon, Dslr, Webcam
  • Face Datasets: PIE, YaleB

No new datasets were introduced; the focus was on evaluating the attack on existing, diverse datasets.

Key Insights and Ideas

  • EC is Vulnerable: The paper convincingly demonstrates that ensemble clustering, despite its robustness, is susceptible to adversarial attacks.
  • Key BPs Matter: Not all basic partitions (BPs) contribute equally to the ensemble's performance. Identifying and attacking the most "vulnerable" or "key" BPs can lead to a more effective and subtle attack.
  • Generative Mimicry is Effective: Using a generative model (VAE) to mimic adversarial attacks allows for controlled, less noticeable perturbations that still significantly degrade clustering performance.
  • Graph Mutation is Crucial: The graph mutation loss is a key innovation, directly targeting the co-association matrix to ensure the adversarial impact is felt in the final clustering, even with subtle BP changes.
  • Inductive Attack Capability: GA2EC's ability to perform attacks in an inductive setting (on unseen data) is a significant advancement, making it more practical for real-world scenarios where the attacker might not have access to the training data.

Key Results

  • Effective Attack: GA2EC consistently degrades clustering performance (measured by Accuracy and NMI) across eleven diverse datasets compared to clean clustering and plain VAE reconstruction.
  • Subtlety: The proposed method achieves a good balance between attack effectiveness and perturbation subtlety, as evidenced by lower Frobenius norms (l2 norms) between original and perturbed BPs compared to standard FGSM attacks (Table 5).
  • Performance Comparison: While the gradient attack with pseudo-labels (GA) sometimes achieves slightly higher degradation, GA2EC offers a more controlled and less noticeable attack, with performance often close to GA.
  • Inductive Setting Success: The method shows consistent performance on unseen data in the inductive setting, demonstrating its generalization capabilities (Figure 6).
  • Ablation Study: The ablation study confirms the importance of each component of the GA2EC framework, particularly the graph mutation loss, for achieving effective and balanced attacks.

Limitations and Future Work

  • Computational Cost: Training the generative model can be time-consuming, with large datasets taking several hours.
  • Parameter Tuning: The balancing parameters ($\lambda_1, \lambda_2, \lambda_3$) require careful tuning to achieve the desired trade-off between attack strength and subtlety.
  • Future Work: The authors suggest exploring adversarial attacks from a Bayesian perspective to gain probabilistic insights and more interpretable solutions. They also mention the potential for extending their work to other types of adversarial attacks and unsupervised learning tasks.

Summary updated

BibTeX
@inproceedings{kumar2022generative,
  author = {Chetan Kumar and Deepak Kumar and Ming Shao},
  title = {{Generative Adversarial Attack on Ensemble Clustering}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
}
Figure