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

Authors: Haoqi He, Xiaokai Lin, Jiancai Chen, Yan Xiao

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results using multiple quantum simulation libraries show that Q-Detection effectively defends against label manipulation and backdoor attacks. The metrics demonstrate that Q-Detection consistently outperforms the baseline methods and is comparable to the state-of-the-art. Theoretical analysis shows that Q-Detection is expected to achieve more than a 20% speedup using quantum computing power.
Researcher Affiliation Academia School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University EMAIL, EMAIL
Pseudocode No The paper describes the methodology, including the bilevel optimization process, QUBO model, and training rules, in textual form. It does not contain a formally structured pseudocode or algorithm block.
Open Source Code Yes The demonstration experiment code can be found at https://github.com/cats1520cakes/Q-Detection-A-Quantum Classical-Hybrid-Poisoning-Attack-Detection-Method.
Open Datasets Yes The experiments are based on the GTSRB traffic sign image dataset, employing various common poisoning attack methods such as Targeted Label Flipping attacks, Bad Nets attacks, and Narcissus backdoor attacks [Jha et al., 2023; Zeng et al., 2023b; Gu et al., 2019] to evaluate the performance of Q-Detection.
Dataset Splits Yes GTSRB is widely used for traffic sign recognition, containing images of 43 categories of traffic signs, with 39,209 training images and 12,630 test images. We used Res Net-18 as the base model for training on the GTSRB dataset.
Hardware Specification Yes All experiments were conducted on a computer with an Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz with 64 GB RAM and an NVIDIA Ge Force RTX 3090 24GB GPU.
Software Dependencies No The paper mentions software like Kaiwu SDK, D-Wave, Qiskit Python API, PyTorch framework, and CUDA, but does not provide specific version numbers for these components.
Experiment Setup Yes We set the batch size to 179, requiring each Q-WAN to train for 220 epochs. Notably, using simulated QA can achieve an average training speed of 1 minute and 10 seconds, which is already very close to the CUDA-accelerated Meta-Sift training time. This suggests promising direct acceleration effects on real Quantum Annealer hardware. When we set the batch size to approximately one-quarter of the original size, i.e., 44, each Weight-Assigning Network needs to train for 892 epochs, resulting in an overall training time of 69 minutes.