強化聯合量子自動編碼器的穩健性

Abstract

量子計算是近年來科學的里程碑，不但引起相當大的關注且廣泛應用於許多領域。然而，目前的噪聲中規模量子裝置（NISQ）容易受到周遭環境、缺陷結構甚至蓄意攻擊的干擾。雖然一些研究表明噪音可以增強微弱信號並保護數據隱私，但我們著重於消除攻擊者刻意製造的噪音以防止意外事件發生。 我們實作三種不同類型的量子自動編碼器。第一種類似於傳統程序將數據的維度降低，然後重新恢復回原始的大小。第二種利用額外的量子位元建立更加彈性的空間使模型能夠探索、發現並有效利用數據中隱藏的關聯性。第三種則是對生成的向量進行額外的操作。一般而言，降低數據的維度可以比擬為壓縮的過程，因為在經過充分訓練後，模型可以藉其完全恢復至原先的狀態。此外，模型也可以檢測並去除數據中存在的噪音或誤導模型的擾動而使其回復至無噪音的狀態。 我們使用結構相似性指數(SSIM)和量子支持向量機(QSVM)評估三種量子自動編碼器的表現和性能並展示其可以藉由受干擾和帶有噪音的數據以進行正確的分類與防止惡意的攻擊。

Quantum computing, a landmark in modern science, has garnered significant attention and is widely used for many applications. However, current Noisy Intermediate-Scale Quantum (NISQ) devices are susceptible to noise from the surrounding environment, defective structure and even deliberate attacks. Although some research has revealed that noise can enhance weak signals and protect the privacy of data, our focus is on eliminating noise crafted by attackers to prevent unexpected events. We conducted an investigation and implemented three different types of quantum autoencoders. The first type follows the conventional procedure of reducing the dimension of the data and then reconstructing it back to its original size. The second type utilizes additional qubits to provide a larger latent space, allowing the model to explore and discover implicit relationships hidden within the data. The third type performs additional operations on the generated relationships. Typically, the data with reduced dimensions can be treated as a compressed state since the model can perfectly recover them after being well-trained. Moreover, if the noise or some perturbations that aim to mislead the model are present in the data, they can be detected and subsequently removed, leaving the state noiseless. We then use structural similarity index measure (SSIM) value and the quantum support vector machines (QSVMs) to evaluate the performance and the robustness of the proposed quantum autoencoders, demonstrating that the perturbed and noised data can now be correctly classified, preventing the model from malicious attacks.