synthetic data and augmentation

This paper investigates whether privacy-preserving data synthesis (PPDS) can safely replace real data for training classifiers, focusing on the fundamental utility–privacy trade-off. To this end, we propose the first end-to-end evaluation framework—encompassing generation, sampling, and classification—that systematically benchmarks state-of-the-art generative models (e.g., GANs and diffusion models) and uniformly quantifies privacy risk via model-agnostic membership inference attacks (MIAs) across diverse benchmark scenarios. Our key contributions are threefold: (1) We empirically uncover intrinsic utility–privacy trade-off patterns across generative architectures; (2) We rigorously assess the efficacy limits of common privacy-mitigation strategies (e.g., differential privacy and output perturbation); and (3) We deliver actionable, scenario-specific guidelines for data publishers—identifying which synthetic data configurations enable safe, utility-preserving substitution of real training data under defined privacy constraints.

This work proposes a GAN-inspired privacy-preserving synthetic data generation method that avoids direct access to original data during training. Instead, it leverages fuzz testing to produce candidate samples and iteratively refines them through a discriminator-guided feedback loop combined with statistical distribution constraints to approximate the original data distribution. By innovatively integrating fuzz testing, adversarial discrimination, and indirect constraint mechanisms, the approach achieves strong privacy guarantees—effectively resisting membership inference and data reconstruction attacks—while preserving high data utility. Extensive experiments on four benchmark datasets demonstrate that the proposed method strikes a superior balance between privacy protection and data fidelity compared to existing techniques.

Existing non-differentially private synthetic data paradigms—core-set selection, dataset distillation, data-free knowledge distillation, and diffusion-based generation—lack rigorous, cross-paradigm evaluation of their actual privacy protection against real-world inference attacks. Method: We conduct a unified empirical assessment using standardized membership inference and data reconstruction attacks, enabling the first horizontal comparison of privacy leakage across these four paradigms. Contribution/Results: Our experiments reveal pervasive and significant privacy leakage across all methods, with diffusion models exhibiting particularly poor resilience under strong adversarial settings. Empirical privacy claims frequently engender false confidence. To address this gap, we propose the first multi-paradigm privacy evaluation framework for synthetic data, advocating formal privacy guarantees over heuristic assurances. This work provides both theoretical grounding and practical warnings for the trustworthy deployment of synthetic data in sensitive applications.

This work addresses the challenge of efficiently generating high-quality training data required for supervised learning in text-to-image generation models. We propose the Guided Adversarial Prompts (GAP) framework—a closed-loop data generation system integrating three core mechanisms: (1) adversarial prompt optimization guided by supervised model loss, (2) target distribution alignment via feature matching or discriminator-based guidance, and (3) online feedback adaptation. GAP is the first method to synergistically couple adversarial generation with explicit distributional constraints, shifting data synthesis from open-loop, static prompting to closed-loop, adaptive refinement. Empirical evaluation across diverse settings—including multi-task learning, heterogeneous model architectures, and distribution shifts (e.g., spurious correlations, unseen domains)—demonstrates substantial improvements in downstream model generalization. Data utilization efficiency increases by up to 3.2× compared to baseline approaches.

This study addresses a critical gap in synthetic data generation (SDG) research, which has predominantly focused on privacy attacks initiated by data recipients while overlooking internal adversaries—such as data owners or generators—who may degrade data quality by perturbing real data. The work formally introduces this internal threat model and proposes targeted perturbation strategies based on label flipping and feature importance manipulation. Through systematic evaluation across multiple mainstream SDG frameworks, the experiments demonstrate that even minimal perturbations can substantially impair downstream task performance and amplify statistical distributional biases. These findings reveal a pronounced vulnerability in current SDG pipelines regarding data integrity and underscore the urgent need for robustness and integrity verification mechanisms in synthetic data workflows.

This work addresses the distribution shift between pre-trained models and real-world deployment environments by proposing a three-stage automated pipeline that leverages diffusion models to generate high-quality, domain-specific synthetic datasets. The approach begins with controlled image inpainting to synthesize objects within target-domain backgrounds, followed by multimodal quality assessment—encompassing object detection accuracy, aesthetic scoring, and vision-language alignment—to evaluate the fidelity of generated samples. Finally, a user preference classifier is introduced to filter outputs according to subjective human standards. To the best of our knowledge, this is the first method to integrate controlled image inpainting, multimodal evaluation, and user preference modeling into an end-to-end framework, substantially reducing reliance on large-scale real-world data collection while efficiently producing deployment-ready domain datasets.

This study systematically evaluates the suitability and effectiveness of synthetic data across three canonical scenarios: data sharing, model training augmentation, and variance reduction in statistical estimation. By integrating formal modeling, theoretical analysis of generative models, and empirical case studies, the work presents the first comprehensive taxonomy of synthetic data applications and delineates their boundaries of applicability. The research elucidates both the potential and fundamental limitations of synthetic data in enhancing privacy preservation, model performance, and statistical stability. It further demonstrates that many existing or proposed use cases are misaligned with the intrinsic properties of synthetic data, thereby providing decision-makers with a principled theoretical framework to assess whether synthetic data is appropriate for addressing specific data availability challenges.

This study addresses the challenge of balancing fidelity, privacy, and downstream utility in synthetic image generation under data-scarce and privacy-sensitive conditions. The authors propose the first unified three-dimensional evaluation framework to systematically compare variational autoencoders (VAEs), generative adversarial networks (GANs), and denoising diffusion probabilistic models (DDPMs) on MNIST, OCTMNIST, and OrganAMNIST benchmarks, incorporating differential privacy mechanisms to assess their impact. The findings reveal that GANs and DDPMs maintain high fidelity and practical utility even under strong privacy constraints, whereas VAE performance degrades significantly. These results highlight fundamental differences in how generative models behave under privacy-preserving conditions and provide empirical guidance for model selection in privacy-sensitive applications.

This study addresses the challenge of balancing privacy preservation and data utility in synthetic data generation for highly regulated financial tabular datasets characterized by severe class imbalance and mixed data types. The authors systematically evaluate the privacy–utility trade-offs of autoencoders, generative adversarial networks (GANs), diffusion models, and Copula-based methods. To better accommodate the unique characteristics of financial data, they propose novel privacy-enhanced GAN and autoencoder architectures. Their experiments provide the first empirical comparison of these generative models in such settings, revealing significant performance differences and offering both methodological innovations and empirical evidence to support the co-optimization of privacy and utility in synthetic financial data generation under stringent regulatory constraints.

Facing data exhaustion in model training and regulatory constraints on private data usage, this paper proposes the Generative Data Refinement (GDR) framework—the first to leverage pretrained generative models for *conditional synthetic data transformation*. GDR simultaneously achieves web-scale data detoxification, anonymization, and distribution preservation without manual prompt engineering. It safely converts private datasets containing sensitive or harmful content into high-quality, privacy-compliant training corpora while preserving semantic diversity and statistical fidelity. Experimental results demonstrate that GDR significantly outperforms state-of-the-art industrial baselines on both anonymization and detoxification tasks: it successfully purifies high-risk datasets and improves downstream model performance across multiple benchmarks. By enabling scalable, privacy-preserving data curation, GDR establishes a new paradigm for sustainable large language model training—one that reconciles dataset expansion with stringent privacy and safety requirements.

Existing single-sample membership inference (MI) methods fail on large-scale, heterogeneous audio datasets, hindering copyright protection for generative audio models. Method: We propose the first verifiable, audio-domain-specific training data attribution framework—Dataset Inference (DI)—a novel set-level inference paradigm that aggregates multi-artist audio samples to determine whether a given dataset contributed to model training. Our approach jointly leverages gradient and output-statistical features from diffusion and autoregressive audio models, incorporates a multi-sample evidence aggregation mechanism, and integrates contrastive benchmark modeling with statistical significance testing. Contribution/Results: Evaluated on multiple open-source large audio models, DI achieves high inference accuracy (AUC > 0.92), substantially outperforming state-of-the-art MI methods. This work provides the first empirically validated, technically feasible solution for audio content copyright auditing and training-data accountability.