Selected work

Representative Papers

A Survey on Generative Modeling with Limited Data, Few Shots, and Zero Shot

Jul 26, 2023arXiv.org

To address the limitation of conventional generative models (e.g., GANs, diffusion models) — their reliance on large-scale labeled data — in data-scarce domains such as medical imaging and remote sensing, this paper proposes a unified framework termed “Generative Modeling under Data Constraints” (GM-DC). We systematically establish a two-dimensional taxonomy: (i) task dimension—encompassing low-data, few-shot, and zero-shot settings; and (ii) methodological dimension—integrating transfer learning, meta-learning, prompt engineering, and multi-paradigm fusion. This work is the first to uncover cross-paradigm adaptation principles and synergistic mechanisms under data constraints. The survey comprehensively analyzes lightweight designs and knowledge transfer strategies for mainstream architectures—including VAEs, GANs, and diffusion models—and identifies critical research gaps while charting emerging trends. As the inaugural holistic GM-DC survey, it is accompanied by an open-source platform for continuous resource updates, providing both theoretical foundations and practical guidance for data-efficient generative modeling.

28 citationsRead paper

Auditing Apple's DifferentialPrivacy.framework: Implementation Bugs, Misconfigurations, and Practical Risks

May 20, 2026

This study presents the first systematic client-side audit of Apple’s DifferentialPrivacy.framework in macOS, addressing concerns about the opacity of its closed-source implementation despite claims of differential privacy (DP) protection. Combining binary reverse engineering, Objective-C interface reconstruction, and runtime testing, the authors evaluate the actual privacy guarantees provided in practice. Their analysis reveals that 5 out of 9 deployed DP mechanisms violate theoretical DP assurances, affecting 87% of data collection tasks in macOS Sonoma and 68% in Sequoia. Critical vulnerabilities include flawed floating-point noise generation and misconfigured local DP parameters. Furthermore, by analyzing publicly available iPhone logs, the study successfully reconstructs sensitive user data—such as Safari-browsed domains and keyboard emoji usage—demonstrating tangible privacy risks stemming from these implementation flaws.

0 citationsRead paper

Towards Irreversible Machine Unlearning for Diffusion Models

Dec 03, 2025

Diffusion model unlearning techniques face reversibility risks—existing fine-tuning-based unlearning methods are vulnerable to the “Diffusion Model Relearning Attack” (DiMRA), enabling reconstruction of supposedly forgotten sensitive data. Method: This work first identifies and formalizes this security vulnerability, introducing the DiMRA framework for empirical validation. To counteract such attacks, we propose DiMUM, an irreversible unlearning method that reprograms internal model representations via semantically consistent memory-replacement data, preserving generation quality while ensuring robust unlearning. Contribution/Results: Experiments demonstrate that DiMRA successfully recovers protected content across multiple state-of-the-art unlearning methods. In contrast, DiMUM achieves >99.2% unlearning success rate and <0.8 FID degradation on benchmarks including CIFAR-10 and CelebA, significantly enhancing resilience against relearning attacks. Our approach establishes a new paradigm for secure and controllable generative AI.

0 citationsRead paper

Instruction Tuning of Large Language Models for Tabular Data Generation-in One Day

Nov 28, 2025

Prior work on table-related tasks focuses predominantly on table question answering and reasoning, largely neglecting table generation—and typically relies on large-scale computational resources and data. Method: This paper pioneers instruction-tuning of large language models (LLMs) for table generation under low-resource constraints: only 7K high-quality instructions, a single A100 GPU, and six hours of training time. We introduce the first high-fidelity, structure-aware instruction dataset for table generation and fine-tune Llama3.1-8B-Instruct to explicitly model header semantics, row-column relationships, and inter-cell semantic consistency. Contribution/Results: Our approach achieves performance on par with GPT-4o across multiple benchmarks while reducing training cost by over 90%. It establishes a new paradigm for lightweight, high-fidelity table generation—demonstrating that effective structural modeling need not require massive resources.

0 citationsRead paper

Generating Synthetic Data with Formal Privacy Guarantees: State of the Art and the Road Ahead

Mar 26, 2025

This work addresses the inherent tension between formal privacy guarantees—particularly differential privacy (DP)—and downstream utility in high-stakes domains such as healthcare and finance. We conduct a systematic empirical evaluation of DP-integrated generative models, including GANs, VAEs, and LLMs. We propose the first domain-specialized, multimodal (tabular/image/text) evaluation framework that jointly quantifies privacy protection and task-specific utility. Our analysis uncovers a substantial performance gap between standard benchmarks and real-world deployment scenarios. Empirical results demonstrate a sharp utility degradation across mainstream methods when ε ≤ 4, revealing a critical misalignment between theoretical privacy guarantees and practical information leakage. The study establishes a reproducible, empirically grounded evaluation paradigm and calibration methodology for privacy-enhancing AI systems.

0 citationsRead paper

Recent publications

Latest Papers

Auditing Apple's DifferentialPrivacy.framework: Implementation Bugs, Misconfigurations, and Practical Risks

May 20, 2026

This study presents the first systematic client-side audit of Apple’s DifferentialPrivacy.framework in macOS, addressing concerns about the opacity of its closed-source implementation despite claims of differential privacy (DP) protection. Combining binary reverse engineering, Objective-C interface reconstruction, and runtime testing, the authors evaluate the actual privacy guarantees provided in practice. Their analysis reveals that 5 out of 9 deployed DP mechanisms violate theoretical DP assurances, affecting 87% of data collection tasks in macOS Sonoma and 68% in Sequoia. Critical vulnerabilities include flawed floating-point noise generation and misconfigured local DP parameters. Furthermore, by analyzing publicly available iPhone logs, the study successfully reconstructs sensitive user data—such as Safari-browsed domains and keyboard emoji usage—demonstrating tangible privacy risks stemming from these implementation flaws.

0 citationsRead paper

Towards Irreversible Machine Unlearning for Diffusion Models

Dec 03, 2025

Diffusion model unlearning techniques face reversibility risks—existing fine-tuning-based unlearning methods are vulnerable to the “Diffusion Model Relearning Attack” (DiMRA), enabling reconstruction of supposedly forgotten sensitive data. Method: This work first identifies and formalizes this security vulnerability, introducing the DiMRA framework for empirical validation. To counteract such attacks, we propose DiMUM, an irreversible unlearning method that reprograms internal model representations via semantically consistent memory-replacement data, preserving generation quality while ensuring robust unlearning. Contribution/Results: Experiments demonstrate that DiMRA successfully recovers protected content across multiple state-of-the-art unlearning methods. In contrast, DiMUM achieves >99.2% unlearning success rate and <0.8 FID degradation on benchmarks including CIFAR-10 and CelebA, significantly enhancing resilience against relearning attacks. Our approach establishes a new paradigm for secure and controllable generative AI.

0 citationsRead paper

Instruction Tuning of Large Language Models for Tabular Data Generation-in One Day

Nov 28, 2025

Prior work on table-related tasks focuses predominantly on table question answering and reasoning, largely neglecting table generation—and typically relies on large-scale computational resources and data. Method: This paper pioneers instruction-tuning of large language models (LLMs) for table generation under low-resource constraints: only 7K high-quality instructions, a single A100 GPU, and six hours of training time. We introduce the first high-fidelity, structure-aware instruction dataset for table generation and fine-tune Llama3.1-8B-Instruct to explicitly model header semantics, row-column relationships, and inter-cell semantic consistency. Contribution/Results: Our approach achieves performance on par with GPT-4o across multiple benchmarks while reducing training cost by over 90%. It establishes a new paradigm for lightweight, high-fidelity table generation—demonstrating that effective structural modeling need not require massive resources.

0 citationsRead paper

Generating Synthetic Data with Formal Privacy Guarantees: State of the Art and the Road Ahead

Mar 26, 2025

This work addresses the inherent tension between formal privacy guarantees—particularly differential privacy (DP)—and downstream utility in high-stakes domains such as healthcare and finance. We conduct a systematic empirical evaluation of DP-integrated generative models, including GANs, VAEs, and LLMs. We propose the first domain-specialized, multimodal (tabular/image/text) evaluation framework that jointly quantifies privacy protection and task-specific utility. Our analysis uncovers a substantial performance gap between standard benchmarks and real-world deployment scenarios. Empirical results demonstrate a sharp utility degradation across mainstream methods when ε ≤ 4, revealing a critical misalignment between theoretical privacy guarantees and practical information leakage. The study establishes a reproducible, empirically grounded evaluation paradigm and calibration methodology for privacy-enhancing AI systems.

0 citationsRead paper

TabTreeFormer: Tree Augmented Tabular Data Generation using Transformers

Jan 02, 2025

Existing Transformer-based approaches for tabular data generation lack domain-specific priors and suffer from poor scalability and low computational efficiency. To address these limitations, we propose a tree-enhanced hybrid architecture coupled with a dual-quantization tokenizer. Our method pioneers the integration of decision trees with Transformers to explicitly capture the non-smoothness and low pairwise correlations inherent in tabular data. The dual-quantization tokenizer jointly optimizes numerical distribution modeling and sequence compression, substantially reducing vocabulary size and sequence length. Key innovations include discretization-aware modeling, non-rotation-invariant constraints, and lightweight sequence encoding. Evaluated on ten benchmark datasets, our approach achieves a 40% improvement in utility over state-of-the-art models, compresses model size to 1/16, and significantly enhances generation fidelity, practical usability, privacy preservation, and inference efficiency.

0 citationsRead paper