Privacy-preserving Machine Learning
Recent studies have indicated the presence of privacy risks associated with the utilization of training data in machine learning models. Empirical evidence from investigations into privacy attacks corroborates these findings. For instance, adversaries can employ membership inference attacks to ascertain whether a data point is part of a training dataset. Another form of attack is data extraction, which involves re-identifying anonymized users and extracting training features such as names, addresses, and phone numbers. Anonymization, which involves the removal of all sensitive information from the original data, is a commonly adopted strategy for safeguarding data privacy. However, research has demonstrated that this heuristic approach remains susceptible to privacy attacks. Consequently, the development of privacy-preserving mechanisms with theoretical guarantees has emerged as a shared research objective among academics and industry professionals. One approach to achieving this objective involves the application of traditional encryption techniques such as multi-party computation (MPC), homomorphic encryption (HE), and trusted execution environments (TEE). Nonetheless, these methods typically necessitate substantial computational and communication complexity, rendering them ill-suited to meet the demands of large-scale machine learning models and voluminous data. To address this challenge, our laboratory concentrates on two cutting-edge machine learning privacy protection mechanisms: Differential Privacy and Federated Learning. Our research encompasses both fundamental theory (privacy-preserving machine learning theory) and system frameworks (privacy-preserving deep learning).
Differentially Private Machine Learning (DPML): Within the framework of differential privacy, our laboratory aims to concentrate on the following aspects: (1) the development of efficient optimization methods for deep learning; (2) the design of deep learning frameworks capable of withstanding various security attacks while preserving privacy; (3) the proposal of differential privacy machine learning systems tailored to different data types (e.g., graph data); (4) an examination of the theoretical limits of fundamental machine learning problems and statistical models under various differential privacy sub-models.
Federated Learning (FL): With respect to federated learning, our current research focuses on the following areas: (1) federated learning with low wireless communication energy consumption; (2) multi-modal federated learning.
Current Member: Junxiao Wang, Zihang Xiang, Liyang Zhu
In the current digital age, machine learning technology has emerged as a core driving force across numerous industries, encompassing fields such as financial services and healthcare. However, the sensitive data processed by these technologies pertains to personal privacy and confidential information, necessitating proper protection. To this end, we have developed several privacy-preserving mechanisms, including differential privacy and homomorphic encryption. Despite these efforts, these technologies are not entirely effective in preventing data misuse and leakage in practice. Moreover, with the rapid development of Artificial Intelligence Generated Content (AIGC), preventing these generative algorithms from producing harmful content has become an urgent issue for academia, governments, and industry. In this context, Machine Unlearning has emerged as a research hotspot. Machine Unlearning refers to the use of algorithms and techniques to delete learned data or specified information from machine learning models in order to avoid generating harmful content and comply with data confidentiality and privacy regulations. Accordingly, our research aims to design and construct fast, effective, and theoretically sound Machine Unlearning algorithms and systems for different application scenarios.
Current Member: Junxiao Wang, Cheng-Long Wang, Liangyu Wang
Faithful and Interpretable AI
Machine learning models are increasingly being applied in critical decision-making scenarios such as healthcare and finance, where transparency and interpretability are essential for establishing trust and ensuring fairness. Explainable Artificial Intelligence (XAI) methods aim to provide insights into how models make predictions. However, current techniques often involve a trade-off between accuracy and interpretability or produce explanations that are difficult to comprehend. This trade-off between accuracy and interpretability limits the effectiveness of XAI methods in practical applications. Furthermore, even when models are interpretable, their explanations may not faithfully reflect the underlying mechanisms of the machine learning model, resulting in unreliability, mistrust, and misunderstanding. To address this issue, our research aims to develop more reliable XAI methods that provide both interpretable and accurate explanations while maintaining the faithfulness of the underlying machine learning model.
Current Member: Lijie Hu
Optoelectronic Acceleration for Machine Learning Algorithms
Due to performance bottlenecks in computing speed and power consumption, traditional electronic computing hardware is unable to meet the needs of artificial intelligence services. Fortunately, due to the ultra-low power consumption, high computing speed, and large-scale parallel capability of optoelectronic information, integrated optoelectronic processors based optical computing is becoming a promising solution for future artificial intelligence hardware. Typical optoelectronic hardware systems use photonic elements for linear multiplication and interconnection, while nonlinear functions and feedback control are handled electronically. In recent years, people have studied a wide range of neural network architectures through photoelectric hardware acceleration, including feedforward neural network (FNN) and convolutional neural network (CNN). At present, the laboratory is working closely with other professor teams to propose neural networks suitable for optoelectronic processors, design optoelectronic systems and photonic chips suitable for neural networks, and send the designed photonic chips to a 180nm standard CMOS process line for physical manufacturing.
Current Member: Cheng-Long Wang, Zihang Xiang
Artificial Intelligence and Optical Systems
Due to the limited ability of the human naked eye to recognize small color changes, we need optical spectroscopy instruments to understand different color interpretation mechanisms. In recent years, the development of miniaturized, portable, and inexpensive spectrometer systems has become a hot topic in academia and industry, which can achieve many emerging on-site, real-time, and in situ spectroscopic/colorimetric analysis applications. However, due to overly simplified optical design and mechanical limitations of compact architecture, the actual performance of miniaturized systems is much lower than that of their desktop systems. Therefore, there is a great need for strategies to improve the color recognition ability of miniaturized systems. In order to solve this problem, the laboratory works closely with other professor teams to fill the gap of high-performance micro spectrometer through the most advanced nano photonics, artificial intelligence algorithms and optical system engineering.
Current Member: Junxiao Wang, Lijie Hu, Xiaochuan Gou, Liangyu Wang, Muhammad Asif Ali