Shulou(Shulou.com)11/24 Report--

CTOnews.com, December 4, according to the official website of Tsinghua University, the research group of Assistant Professor Deng Dongling of the Cross Information Research Institute of Tsinghua University, in cooperation with Wang Haohua and Song Chao of the School of Physics of Zhejiang University, realized quantum countermeasure machine learning in the superconducting system for the first time.

The cover of Natural Computational Science: the early research on quantum counteraction learning schematics against machine learning can be traced back to the problem of spam filtering (spam filtering), involving the game between the sender and the boycott of spam. Generally speaking, when the user's email address is known to the outside world, some malicious parties may send advertising emails, computer viruses and so on to this mailbox for commercial gain. In order to resist this kind of behavior, people have developed mail filters to distinguish normal mail from malicious mail and block the latter. In order to avoid the detection of the mail filter, the senders of malicious messages will take a series of measures, such as modifying the characteristic words in the malicious messages, adding normal words and so on.

With the development of deep learning, deep learning model has been widely used in face recognition, autopilot and other fields. However, it is found that the deep learning model also has the threat of being attacked by counter samples. In a deep learning model that has been trained to correctly identify pandas, even the addition of a disturbance imperceptible to the naked eye is likely to turn the prediction given by the model into a gibbon. If this kind of attack is not resolved and maliciously exploited, it may lead to serious security risks. For example, in a self-driving car, a safety accident may occur if a stop sign in front is affixed with a well-designed anti-disturbance film and judged by the car's identification program to drive at a normal speed. In machine learning assisted medical diagnosis, if the MRI image is maliciously added with a small disturbance, it may also lead to medical malpractice. Thus it can be seen that the research of antagonistic learning is very important for the security of the practical application of machine learning.

In the field of quantum machine learning, the theoretical work in recent years has shown that under some specific tasks, the quantum classification model and generation model have provable advantages over the classical model and are guaranteed by complexity theory. In the past two years, the concept of quantum countermeasure machine learning has also been proposed and attracted wide attention. However, there are still many challenges in demonstrating the vulnerability and defense of the quantum learning model to counter attacks on current medium-scale noisy (NISQ) quantum devices. In this study, the Deng Dongling team of the Cross Information Research Institute of Tsinghua University designed a learning model to deal with classical data and quantum multi-body data, and worked with the superconducting quantum computing team of Zhejiang University to successfully realize the learning of high-dimensional data, the disclosure of vulnerability against attacks and the display of corresponding defense methods on quantum devices for the first time.

The quantum adversarial learning experiment for classical input data first carries on the adversarial learning of high-dimensional classical data, and uses nuclear magnetic resonance image (MRI) as the training data. In order to achieve higher classification accuracy on the current superconducting quantum chips with noise, the scheme of interleaving embedding of variational parameters and input data is adopted in the experiment. After training to convergence and achieving good results, through the way of generating anti-noise, it is found that the quantum classification model will give wrong classification judgment to the samples with anti-noise, which reveals the vulnerability of the current model to counterattack. In order to enhance the robustness of the model against potential anti-noise, the model is retrained by confrontation training in the experiment. The corresponding results show that after the confrontation training, the confrontation samples that led to the misjudgment of the model can not make the updated model go wrong again, and the defense effect of the confrontation training has been verified.

For the antagonistic sample generation of quantum input data, in addition, the classification of quantum multibody data by quantum learning model and the corresponding countermeasure sample generation are studied experimentally. In the experiment, two kinds of quantum state data are generated by the evolution of many-body Hamiltonian, namely, local state and thermal state. After the training, the model can distinguish these quantum states with nearly 100% accuracy. By generating counternoise, the experiment reveals that even if the antagonistic samples maintain local / thermal properties similar to those of the original samples, these samples can make the model give a wrong classification, thus showing that the model is vulnerable to counternoise. The 36-bit superconducting quantum chip used in the experiment has an extensible nearest neighbor connectivity architecture. Its high programming flexibility and 99.94% hammer 99.4% fidelity single / double-bit quantum gates provide the basis for the experimental implementation of the model and can be used to explore more unknown quantum machine learning architectures.

CTOnews.com learned that the paper "Experimental demonstration of Quantum antagonistic Learning through Programmable Superconducting Quantum bits" (Experimental quantum adversarial learning with programmable superconducting qubits) was recently published in Natural Computational Science (Nature Computational Science) in the form of a cover paper and received a column review in the journal.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.