Bochao Liu

I'm a Ph.D student at Institute of Information Engineering, Chinese Academy of Sciences in Beijing, advised by Prof. Shiming Ge. I received my B.S. degree in Electronical Information Science and Technology from the School of Information Science and Engineering in Shandong University. From December 2023 to the end of May 2024, I interned at the AI Safety Group of Baidu. From June 2024 to the end of August 2024, I interned at the Data Storage Product Line of Huawei.

My research interests include AIGC, generative models, differentially private learning, computer vision. Feel free to contact me via email: liubochao@iie.ac.cn.

Email  /  CV  /  Scholar  /  Github

News

• Our paper: Private Gradient Estimation is Useful for Generative Modeling is accepted as Oral by MM 2024.
• Our paper: Towards Personalized Federated Learning via Comprehensive Knowledge Distillation is accepted by SMC 2024.
• Our paper: Learning Differentially Private Diffusion Models via Stochastic Adversarial Distillation is accepted by ECCV 2024.
• Our paper: Model Conversion via Differentially Private Data-Free Distillation is accepted by IJCAI 2023.
• Our paper: Learning Privacy-Preserving Student Networks via Discriminative-Generative Distillation is accepted by TIP 2023.
• Our paper: Privacy-Preserving Student Learning with Differentially Private Data-Free Distillation is accepted by MMSP 2022 and received Best Student Paper-Honorable Mention Award.

Selected Publications (Full list)

This paper introduces a novel private generative modeling approach that utilizes Hamiltonian dynamics with sliced score matching to achieve differential privacy by adding noise during gradient estimation. The model successfully generates 256×256 resolution data by progressively approximating the private dataset's manifold, and extensive experiments validate its effectiveness.

This paper trains a private diffusion model by a stochastic adversarial distillation method. Specifically, we first train a diffusion model as a teacher and then train a student by distillation, in which we achieve differential privacy by adding noise to the gradients from other models to the student. For better generation quality, we introduce a discriminator to distinguish whether an image is from the teacher or the student, which forms the adversarial training.

This paper propose a learning approach termed differentially private data-free distillation (DPDFD) for model conversion that can convert a pretrained model (teacher) into its privacy-preserving counterpart (student) via an intermediate generator without access to training data. This work implements SOTA in the field of differentially private learning and theoretically proves that DPDFD can guarantee differential privacy and well convergence.

This paper proposed a discriminative-generative distillation approach to learn privacy-preserving deep models. The key idea is taking models as a bridge to distill knowledge from private data and then transfer it to learn a student network via two streams.

This paper presents an effective teacher-student learning approach to train privacy-preserving deep learning models via differentially private data-free distillation. The main idea is generating synthetic data to learn a student that can mimic the ability of a teacher well-trained on private data.

Blogs

Patents