Q-Detection: A Quantum-Classical Hybrid Poisoning Attack Detection Method

Haoqi He, Xiaokai Lin, Jiancai Chen, Yan XiaoIJCAI 2025 · DOI

TL;DR

This paper introduces Q-Detection, a novel quantum-classical hybrid method for detecting data poisoning attacks in image classification. By leveraging quantum computing for a bilevel optimization problem, Q-Detection aims to efficiently filter out malicious data, leading to more robust and accurate deep learning models.

What Problem Does This Paper Address?

In modern machine learning, especially for critical applications like autonomous driving or medical diagnosis, the integrity of training data is paramount. Data poisoning attacks, where malicious actors inject corrupted data into the training set, pose a significant threat. These attacks can degrade model performance or manipulate predictions, leading to severe consequences. While classical defense mechanisms exist, they often struggle with the scale and complexity of modern datasets, leading to high computational costs for detection. This paper tackles the challenge of efficiently and effectively detecting poisoned data in large datasets, proposing a solution that harnesses the potential speedups offered by quantum computing.

Methods

The core of Q-Detection is a quantum-classical hybrid approach that transforms the data poisoning detection problem into a bilevel optimization task. This task is then addressed using a Quantum Weight-Assigning Network (Q-WAN), a one-hidden-layer network trained using quantum computing.

Here's a breakdown of the key methodological components:

  • Bilevel Optimization: Q-Detection frames the problem as a bilevel optimization. The outer loop aims to maximize the loss on potentially poisoned data (Adversarial Filtering Stage), while the inner loop minimizes the loss on clean data (Selective Learning and Actual-update Stages). This iterative process helps the Q-WAN learn to assign weights to data samples, effectively distinguishing between clean and poisoned data.
  • Quantum Weight-Assigning Network (Q-WAN): This is the novel component. The Q-WAN is trained using quantum devices to assign weights to data samples. Its parameters are optimized through a process that involves comparing "free" and "guided" states, leveraging quantum phenomena.
  • QUBO Formulation: The training of the Q-WAN is formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem. This allows for the utilization of various quantum computing devices, including Quantum Annealers, Coherent Quantum Photon Computers, and Gate Model Superconducting Quantum Computers.
  • Hybrid Approach: Q-Detection combines classical deep learning models (like ResNet-18) with quantum computation. The Q-WAN acts as a filter, identifying clean data subsets, which are then used to train a classical model.

Datasets

The experiments were conducted using the GTSRB (German Traffic Sign Recognition Benchmark) dataset. This dataset is widely used for traffic sign recognition and contains images across 43 categories. The paper notes that the experiments involved poisoning this dataset with various attack types.

Key Insights and Ideas

  • Quantum Advantage for Data Filtering: The paper demonstrates that quantum computing can offer a significant advantage in the computationally intensive task of filtering poisoned data, a crucial step in robust machine learning.
  • Q-WAN as a Novel Detector: The introduction of the Q-WAN, trained via quantum optimization, provides a new paradigm for data poisoning detection, moving beyond purely classical anomaly detection methods.
  • Bilevel Optimization for Robustness: Framing the detection as a bilevel optimization problem allows for a more nuanced approach, actively seeking to maximize errors on poisoned data while minimizing them on clean data.
  • Scalability with Qubits: The performance of Q-Detection, particularly its ability to achieve SOTA results, is shown to be highly dependent on the number of available qubits. This suggests a clear path for future improvements as quantum hardware scales.

Key Results

  • SOTA Performance: Q-Detection consistently outperforms baseline methods and achieves state-of-the-art (SOTA) results, particularly when using a larger number of qubits (e.g., 5000 qubits).
  • Effective Filtering: The filtered subsets of data used for training new models resulted in accuracy surpassing baseline methods. For instance, in Figure 1, Q-Detection showed improved overall and specific label accuracy under targeted label-flipping attacks.
  • Speedup Potential: Theoretical analysis suggests Q-Detection is expected to achieve more than a 20% speedup using quantum computing power compared to classical methods like Meta-Sift. Simulated experiments showed that Q-Detection with integrated CIM computing power could be approximately 2.17 minutes, a significant improvement over Meta-Sift's 3 minutes.
  • Qubit Dependency: The effectiveness of Q-Detection is directly correlated with the number of qubits. While smaller qubit counts (e.g., 20 neurons/qubits) showed poor performance, larger configurations (500 and especially 5000 qubits) demonstrated superior filtering capabilities.

Limitations and Future Work

  • Hardware Dependency: The current experiments rely heavily on simulated quantum computing due to the limitations of available quantum hardware. The authors acknowledge that real-world acceleration is expected with the advancement of quantum computing technology and the gradual increase in qubits.
  • Qubit Limitations in Simulations: For certain quantum devices like CQPC, simulations were limited to 600 qubits, impacting the performance comparison at higher qubit counts.
  • Future Work: The paper highlights the potential for Q-Detection to achieve even greater acceleration on real quantum hardware. Future work will likely focus on implementing and testing Q-Detection on actual quantum computers to fully realize its speedup potential and further explore its effectiveness across a wider range of poisoning attacks and datasets. The authors also suggest that the Q-WAN's architecture and training could be further optimized.

Summary updated

BibTeX
@inproceedings{he2025q,
  author = {Haoqi He and Xiaokai Lin and Jiancai Chen and Yan Xiao},
  title = {{Q-Detection: A Quantum-Classical Hybrid Poisoning Attack Detection Method}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/593},
}
Figure