Zihao Zhao

I am currently a third-year Ph.D. student in the School of Computer Science and Engineering at Georgia Institute of Technology. I am very fortunate to be advised by Kai Wang.

My research interests lie primarily in machine learning, optimization, and AI for social impact. Currently, I mainly study on topics related to bilevel optimization and decision-focused learning.

Feel free to email me at zzhao628@gatech.edu if you'd like to discuss my research or explore a potential collaboration.

Email / CV / Google Scholar / LinkedIn / Github / Blogs

Works during PhD

Preprint:
	Variance-Optimal Control Variates for Learning with Black-box Feedback Zihao Zhao, Shuhan Zhang, Kai Wang arXiv preprint, 2026. Problem: When learning from black-box oracles (humans, solvers, external tools), a learned action-value function is the standard control variate. However, even an exact action-value function can be far from variance-optimal, because actions also affect the gradient estimator through shared parameters. Our solution: (1) A simple unbiased correction that requires no extra oracle calls and can reduce variance by an arbitrarily large factor; and (2) an exact per-action decomposition of the residual variance, which yields a closed-form variance-minimizing correction for neural-network parameters via a simple projection.
Publications:
	Decision-Focused Learning in MDPs: An Occupancy Measure Approach [Code] Zihao Zhao, Ashwath K. Karunakaram, Ali Eshragh, Yuexing Li, Kai Wang Neural Information Processing Systems (NeurIPS), 2026. Problem: Decision-focused learning for sequential decisions is hard because (1) predicted parameters enter the constraints, biasing the objective away from the true environment; and (2) computational cost scales with the size of the state and action spaces. Our solution: (1) An augmented Lagrangian loss that provably recovers the true regret minimizer; and (2) end-to-end learning that aggregates similar states into "super-states," so computation scales only with the number of super-states.
	A Fully First-Order Layer for Differentiable Optimization [Code] Zihao Zhao, Kai-Chia Mo, Shing-Hei Ho, Brandon Amos, Kai Wang International Conference on Machine Learning (ICML), Spotlight (top 2.2%), 2026. We introduce an active-set Lagrangian hypergradient oracle that avoids Hessian evaluations and provides finite-time, non-asymptotic approximation guarantees. We show that an approximate hypergradient can be computed using only first-order information in $\tilde{\mathcal{O}}(1)$ time, leading to an overall complexity of $\tilde{\mathcal{O}}(\delta^{-1}\epsilon^{-3})$ for constrained bilevel optimization, which matches the best known rate for non-smooth non-convex optimization.
	Diffusion-DFL: Decision-focused Diffusion Models for Stochastic Optimization [Code] Zihao Zhao, Christopher Yeh, Lingkai Kong, Kai Wang International Conference on Learning Representations (ICLR), 2026. We propose the first diffusion-based DFL approach for stochastic optimization, which trains a diffusion model to capture complex uncertainty in problem parameters. We develop two end-to-end training techniques to integrate the diffusion model into decision-making: reparameterization and score function approximation.

Works before PhD

	Federated PAC-Bayesian Learning on Non-IID Data [Code] Zihao Zhao, Yang Liu, Wenbo Ding, Xiao-Ping Zhang Accepted by IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024. We introduce a non-vacuous federated PAC-Bayesian generalization error bound tailored for non-IID local data, and presented an innovative Gibbs-based algorithm for its optimization. Tightness of the bound has been validated by real-world datasets.
	AQUILA: Communication Efficient Federated Learning with Adaptive Quantization in Device Selection Strategy [Code] Zihao Zhao, Yuzhu Mao, Zhenpeng Shi, Yang Liu, Tian Lan, Wenbo Ding, Xiao-Ping Zhang IEEE Transactions on Mobile Computing (TMC), 2023. We craft an optimization problem to minimize the impact of skipped client updates, then derive an optimal quantization precision strategy, demonstrating comparable model performance with a 60.4% communication costs reduction on both IID and non-IID scenarios.
	Inclusive Data Representation in Federated Learning: A Novel Approach Integrating Textual and Visual Prompt [Code] Zihao Zhao, Zhenpeng Shi, Yang Liu, Wenbo Ding Proceedings of the 2023 ACM International Joint Conference on Pervasive and Ubiquitous Computing & the 2023 ACM International Symposium on Wearable Computing (UbiComp-CPD), 2023. Oral presentation, Best Paper Runner-Up Award 🏆. We develop a twin prompt tuning algorithm -- integrating both visual and textual modalities, enhance the data representation, and achieve superior performance over all baseline models in 7 datasets.
	Deep Leakage from Model in Federated Learning [Code] Zihao Zhao, Mengen Luo, Wenbo Ding. Conference on Parsimony and Learning (CPAL), 2024. Oral presentation. Also on IEEE East Asian School of Information Theory (EASIT), 2022. We introduce a model-based attack to recovery privacy data of users using a novel matrix Frobenius norm loss functions, achieving 92% recovery accuracy and 32% higher than gradient-based attacks.
	SAFARI: Sparsity-Enabled Federated Learning with Limited and Unreliable Communications [Code] Yuzhu Mao, Zihao Zhao, Meilin Yang, Le Liang, Yang Liu, Wenbo Ding, Tian Lan, Xiao-Ping Zhang IEEE Transactions on Mobile Computing (TMC), 2023. We develop a sparsity-enabled framework that employs a client similarity matrix to address unreliable communications, ensuring federated learning convergence even with even with 60% weight pruning and 80% client update loss.
	Towards Efficient Communications in Federated Learning: A Contemporary Survey Zihao Zhao, Yuzhu Mao, Yang Liu, Linqi Song, Ye Ouyang, Xinlei Chen, Wenbo Ding Journal of the Franklin Institute, 2023. We comprehensively review the challenges and advancements in communications of federated learning, highlight improvements in communication efficiency, environment, and resource allocation, and forecast future research directions on communication-efficient FL algorithms.
	Communication-Efficient Federated Learning with Adaptive Quantization [Code] Yuzhu Mao, Zihao Zhao, Guangfeng Yan, Yang Liu, Tian Lan, Linqi Song, Wenbo Ding ACM Transactions on Intelligent Systems and Technology (TIST), 2022. We adjust the quantization precision for optimal precision by brute-force searching, allow a 25%-50% decrease in transmission compared to existing methods, and demonstrated resilience to up to 90% client dropout rates.
	MAGLeak: A Learning-based Side-channel Attack for Password Recognition with Multiple Sensors in IIoT Environment [Code] Dajiang Chen, Zihao Zhao, Xue Qin, Yaohua Luo, Mingsheng Cao, Hua Xu, Anfeng Liu IEEE Transactions on Industrial Informatics (TII), 2020. We develop a side-channel-based password recognition system utilizing the 3 types of smartphone sensors for password detection, surpassing previous methods with up to 98% accuracy on limited training data.