arXiv subjects

Haoyang Hu

Publications and source records attributed to Haoyang Hu.

9 recordsLinked to original sources

arXiv2026

Optimal Domain-Aware Privacy Mechanisms for Synthetic Data Generation

Differential privacy (DP) imposes fundamental trade-offs between privacy and statistical fidelity in synthetic data generation. While access to public data has been shown to improve these trade-offs empirically, existing approaches use public data only indirectly, through pre-processing (e.g., using pre-trained generative models) or post-processing steps (e.g., matching target statistics estimated from public datasets), while relying on domain-agnostic DP mechanisms. In this work, we lay the theoretical framework to study the principled incorporation of public data into DP mechanisms themselves. We consider normalized histograms as distribution estimators and characterize the asymptotically optimal domain-aware privacy mechanism within a specific class of DP mechanisms. We introduce PubMix, a public-data-aware DP mechanism that can be used in histogram-based data synthesis pipelines. Our experiments demonstrate that PubMix significantly improves synthetic data generation quality compared to domain-agnostic privacy mechanisms.

cs.IT↗

arXiv2026

Turbulence Physics Governs a Scaling Law for the Machine-Learning Predictability Ceiling in Chaotic Flow

For centuries, the intrinsic chaos of unsteady fluid motion has stood as a formidable barrier to long-term forecasting. While machine learning (ML) has recently emerged as a transformative paradigm for predicting flow evolution, it encounters a pervasive yet unexplained "performance wall": an inevitable deterioration in accuracy as the forecast horizon extends. Here, we demonstrate that this deterioration is not a deficiency of model architecture, no matter how state-of-the-art, but a fundamental constraint imposed by the underlying system, which can be understood through turbulence theory established decades ago. In the setting of bluff body flow, a canonical phenomenon for spatiotemporal complexity in fluid mechanics, we reveal a scaling law governing the deterioration of ML predictability, derived from a Kolmogorov-inspired framework and validated through high-fidelity simulations. Our findings establish a closed loop between the predictability ceiling and its interpretation, bridging the gap between transparent physical theories and modern black-box inference. More broadly, this work provides a theoretical compass for constructing trustworthy ML in complex dynamical systems across the physical sciences.

physics.flu-dyn↗

arXiv2026

Drag within Prior Distribution: Text-Conditioned Point-Based Image Editing within Distribution Constraints

Diffusion-based point editing methods have gained significant traction in image editing tasks due to their ability to manipulate image semantics and fine details by applying localized perturbations on the manifold of noise latent. However, these approaches face several limitations. Traditional point-based editing relies on pairs of handle and target points to define motion trajectories, which can introduce ambiguity or unnecessary alterations. Furthermore, when the distance between the handle and target points is large, the accumulated perturbations often cause the noise latent deviation from inversion score trajectory, resulting in unnatural artifacts. To address these issues in global editing tasks, we introduce a CLIP-based model to evaluate and guide intermediate editing steps, ensuring that the generated results remain both semantically aligned. Additionally, we propose a prior-preservation loss that constrains the optimized latent code to stay within the sampling space of the diffusion prior, improving consistency with the original data distribution, to ensure the model generates images along a familiar score trajectory. For fine-grained tasks, we present a directionally-weighted point tracking mechanism that steers the editing process toward the target direction within similar feature regions. This improves both the tracking accuracy and generation quality, while also reducing the editing time.

cs.CV↗

arXiv2026

Differentially Private Secure Multiplication: Beyond Two Multiplicands

We study the problem of differentially private (DP) secure multiplication in distributed computing systems, focusing on regimes where perfect privacy and perfect accuracy cannot be simultaneously achieved. Specifically, N nodes collaboratively compute the product of M private inputs while guaranteeing epsilon-DP against any collusion of up to T nodes. Prior work has characterized the fundamental privacy-accuracy trade-off for the multiplication of two multiplicands. In this paper, we extend these results to the more general setting of computing the product of an arbitrary number M of multiplicands. We propose a secure multiplication framework based on carefully designed encoding polynomials combined with layered noise injection. The proposed construction generalizes existing schemes and enables the systematic cancellation of lower-order noise terms, leading to improved estimation accuracy. We explore two regimes: (M-1)T+1 <= N <= MT and N = T+1. For (M-1)T+1 <= N <= MT, we characterize the optimal privacy--accuracy trade-off. When N = T+1, we derive nontrivial achievability and converse bounds that are asymptotically tight in the high-privacy regime.

cs.IT↗

arXiv2026

Confundo: Learning to Generate Robust Poison for Practical RAG Systems

Retrieval-augmented generation (RAG) is increasingly deployed in real-world applications, where its reference-grounded design makes outputs appear trustworthy. This trust has spurred research on poisoning attacks that craft malicious content, inject it into knowledge sources, and manipulate RAG responses. However, when evaluated in practical RAG systems, existing attacks suffer from severely degraded effectiveness. This gap stems from two overlooked realities: (i) content is often processed before use, which can fragment the poison and weaken its effect, and (ii) users often do not issue the exact queries anticipated during attack design. These factors can lead practitioners to underestimate risks and develop a false sense of security. To better characterize the threat to practical systems, we present Confundo, a learning-to-poison framework that fine-tunes a large language model as a poison generator to achieve high effectiveness, robustness, and stealthiness. Confundo provides a unified framework supporting multiple attack objectives, demonstrated by manipulating factual correctness, inducing biased opinions, and triggering hallucinations. By addressing these overlooked challenges, Confundo consistently outperforms a wide range of purpose-built attacks across datasets and RAG configurations by large margins, even in the presence of defenses. Beyond exposing vulnerabilities, we also present a defensive use case that protects web content from unauthorized incorporation into RAG systems via scraping, with no impact on user experience.

cs.CR↗

arXiv2025

BERTector: An Intrusion Detection Framework Constructed via Joint-dataset Learning Based on Language Model

Intrusion detection systems (IDS) are widely used to maintain the stability of network environments, but still face restrictions in generalizability due to the heterogeneity of network traffics. In this work, we propose BERTector, a new framework of joint-dataset learning for IDS based on BERT. BERTector integrates three key components: NSS-Tokenizer for traffic-aware semantic tokenization, supervised fine-tuning with a hybrid dataset, and low-rank adaptation for efficient fine-tuning. Experiments show that BERTector achieves state-of-the-art detection accuracy, strong generalizability, and excellent robustness. BERTector achieves the highest accuracy of 99.28% on NSL-KDD and reaches the average 80% detection success rate against four perturbations. These results establish a unified and efficient solution for modern IDS in complex and dynamic network environments.

cs.CR↗

arXiv2025

Vortex-Induced Drag Forecast for Cylinder in Non-uniform Inflow

In this letter, a physics-based data-driven strategy is developed to predict vortex-induced drag on a circular cylinder under non-uniform inflow conditions - a prevalent issue for engineering applications at moderate Reynolds numbers. Traditional pressure-signal-based models exhibit limitations due to complex vortex dynamics coupled with non-uniform inflow. To address this issue, a modified fully connected neural network (FCNN) architecture is established that integrates upstream velocity measurements (serving as an inflow calibration) with pressure-signal-based inputs to enhance predictive capability (R^2 ~ 0 to 0.75). Direct numerical simulations (DNS) at Reynolds number Re = 4000 are implemented for model training and validation. Iterative optimizations are conducted to derive optimized input configurations of pressure sensor placements and velocity components at upstream locations. The optimized model achieves an R^2 score of 0.75 in forecasting high-amplitude drag coefficient fluctuations (C_d=0.2 - 1.2) within a future time window of one time unit. An exponential scaling between model performance and optimized pressure signal inputs is observed, and the predictive capability of sparsely distributed but optimized sensors is interpreted by the scaling. The optimized sensor placements correspond to the physical mechanism that the flow separation dynamics play a governing role in vortex-induced drag generation. This work advances machine learning applications in fluid-structure interaction systems, offering a scalable strategy for forecasting statistics in turbulent flows under real-world engineering conditions.

physics.flu-dyn↗

arXiv2025

Differentially Private Secure Multiplication with Erasures and Adversaries

We consider a private distributed multiplication problem involving N computation nodes and T colluding nodes. Shamir's secret sharing algorithm provides perfect information-theoretic privacy, while requiring an honest majority, i.e., N \ge 2T + 1. Recent work has investigated approximate computation and characterized privacy-accuracy trade-offs for the honest minority setting N \le 2T for real-valued data, quantifying privacy leakage via the differential privacy (DP) framework and accuracy via the mean squared error. However, it does not incorporate the error correction capabilities of Shamir's secret-sharing algorithm. This paper develops a new polynomial-based coding scheme for secure multiplication with an honest minority, and characterizes its achievable privacy-utility tradeoff, showing that the tradeoff can approach the converse bound as closely as desired. Unlike previous schemes, the proposed scheme inherits the capability of the Reed-Solomon (RS) code to tolerate erasures and adversaries. We utilize a modified Berlekamp-Welch algorithm over the real number field to detect adversarial nodes.

cs.IT↗

arXiv2022

Coded Distributed Computing for Hierarchical Multi-task Learning

In this paper, we consider a hierarchical distributed multi-task learning (MTL) system where distributed users wish to jointly learn different models orchestrated by a central server with the help of a layer of multiple relays. Since the users need to download different learning models in the downlink transmission, the distributed MTL suffers more severely from the communication bottleneck compared to the single-task learning system. To address this issue, we propose a coded hierarchical MTL scheme that exploits the connection topology and introduces coding techniques to reduce communication loads. It is shown that the proposed scheme can significantly reduce the communication loads both in the uplink and downlink transmissions between relays and the server. Moreover, we provide information-theoretic lower bounds on the optimal uplink and downlink communication loads, and prove that the gaps between achievable upper bounds and lower bounds are within the minimum number of connected users among all relays. In particular, when the network connection topology can be delicately designed, the proposed scheme can achieve the information-theoretic optimal communication loads. Experiments on real datasets show that our proposed scheme can reduce the overall training time by 17% $\sim$ 26% compared to the conventional uncoded scheme.

cs.IT↗