arXiv subjects

Meisam Mohammady

Publications and source records attributed to Meisam Mohammady.

12 recordsLinked to original sources

arXiv2026

What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks

Large language model (LLM)-powered content moderation systems are a critical defense against harmful online content. However, they operate primarily on tokenized text and often overlook visual cues that humans naturally use when interpreting content. We show that this limitation creates a fundamental vulnerability: content readily recognized as harmful by humans can evade automated moderation. To systematically study this problem, we introduce Human-Perceptible Adversarial Attacks (HPAA), which embed harmful expressions into otherwise benign text using visually salient typographic manipulations. HPAA strategically combines features such as spacing, emphasis, and spatial arrangement to preserve human recognition while reducing machine detectability. Operating in a black-box setting with a small query budget, the attack automatically generates evasive content without model access or gradient information. We evaluate HPAA on multiple datasets and thirteen widely deployed moderation systems, including commercial APIs and state-of-the-art open-source guardrails. With only three detector queries, generated attacks achieve over 86\% human recognition while keeping detection rates below 1\% across evaluated systems. We further identify the typographic factors driving successful evasion, analyze why current moderation architectures fail to capture these signals, and discuss practical defenses. Our findings reveal a fundamental blind spot in current LLM-based moderation systems and motivate moderation approaches that better align with human perceptual understanding.

cs.CR↗

arXiv2026

A Nonlinear Deficiency Identity for the Riemann Zeta Function with Optimal Approximation Rates

We introduce a deficiency-based representation and approximation framework for values of the Riemann zeta function. The method is based on comparing two nonlinear accumulation mechanisms: global transformation of a base partial sum and local transformation of each term. Their gap defines a cumulative deficiency functional that yields the exact identity \[ ζ(q)=ζ(p)^{q/p}-D_{\infty}^{(p,q)}, \qquad q>p>1. \] This converts zeta approximation into estimation of a nonlinear deficit. We derive corrected estimators that remove first-order bias and prove the convergence law \[ B_n^{(p,q)}-ζ(q)=O\!\left(n^{-\min(2p-2,q-1)}\right). \] For odd targets, suitable choices of the base exponent recover the natural truncation rate while preserving the structural identity. Numerical experiments for $ζ(3),ζ(5),ζ(7)$ confirm theory, demonstrate strong finite-sample behavior, and illustrate extension to spectral zeta functions. The contribution is structural rather than replacing classical Euler--Maclaurin methods: we provide a unified nonlinear viewpoint on zeta approximation, convexity-induced correction terms, and tunable approximation families.

math.GM↗

arXiv2026

Lap2: Revisiting Laplace DP-SGD for High Dimensions via Majorization Theory

Differentially Private Stochastic Gradient Descent (DP-SGD) is a cornerstone technique for ensuring privacy in deep learning, widely used in both training from scratch and fine-tuning large-scale language models. While DP-SGD predominantly relies on the Gaussian mechanism, the Laplace mechanism remains underutilized due to its reliance on L1 norm clipping. This constraint severely limits its practicality in high-dimensional models because the L1 norm of an n-dimensional gradient can be up to sqrt(n) times larger than its L2 norm. As a result, the required noise scale grows significantly with model size, leading to poor utility or untrainable models. In this work, we introduce Lap2, a new solution that enables L2 clipping for Laplace DP-SGD while preserving strong privacy guarantees. We overcome the dimensionality-driven clipping barrier by computing coordinate-wise moment bounds and applying majorization theory to construct a tight, data-independent upper bound over the full model. By exploiting the Schur-convexity of the moment accountant function, we aggregate these bounds using a carefully designed majorization set that respects the L2 clipping constraint. This yields a multivariate privacy accountant that scales gracefully with model dimension and enables the use of thousands of moments. Empirical evaluations demonstrate that our approach significantly improves the performance of Laplace DP-SGD, achieving results comparable to or better than Gaussian DP-SGD under strong privacy constraints. For instance, fine-tuning RoBERTa-base (125M parameters) on SST-2 achieves 87.88% accuracy at epsilon=0.54, outperforming Gaussian (87.16%) and standard Laplace (48.97%) under the same budget.

cs.CR↗

arXiv2025

PLRV-O: Advancing Differentially Private Deep Learning via Privacy Loss Random Variable Optimization

Differentially Private Stochastic Gradient Descent (DP-SGD) is a standard method for enforcing privacy in deep learning, typically using the Gaussian mechanism to perturb gradient updates. However, conventional mechanisms such as Gaussian and Laplacian noise are parameterized only by variance or scale. This single degree of freedom ties the magnitude of noise directly to both privacy loss and utility degradation, preventing independent control of these two factors. The problem becomes more pronounced when the number of composition rounds T and batch size B vary across tasks, as these variations induce task-dependent shifts in the privacy-utility trade-off, where small changes in noise parameters can disproportionately affect model accuracy. To address this limitation, we introduce PLRV-O, a framework that defines a broad search space of parameterized DP-SGD noise distributions, where privacy loss moments are tightly characterized yet can be optimized more independently with respect to utility loss. This formulation enables systematic adaptation of noise to task-specific requirements, including (i) model size, (ii) training duration, (iii) batch sampling strategies, and (iv) clipping thresholds under both training and fine-tuning settings. Empirical results demonstrate that PLRV-O substantially improves utility under strict privacy constraints. On CIFAR-10, a fine-tuned ViT achieves 94.03% accuracy at epsilon approximately 0.5, compared to 83.93% with Gaussian noise. On SST-2, RoBERTa-large reaches 92.20% accuracy at epsilon approximately 0.2, versus 50.25% with Gaussian.

cs.CR↗

arXiv2024

Universally Harmonizing Differential Privacy Mechanisms for Federated Learning: Boosting Accuracy and Convergence

Differentially private federated learning (DP-FL) is a promising technique for collaborative model training while ensuring provable privacy for clients. However, optimizing the tradeoff between privacy and accuracy remains a critical challenge. To our best knowledge, we propose the first DP-FL framework (namely UDP-FL), which universally harmonizes any randomization mechanism (e.g., an optimal one) with the Gaussian Moments Accountant (viz. DP-SGD) to significantly boost accuracy and convergence. Specifically, UDP-FL demonstrates enhanced model performance by mitigating the reliance on Gaussian noise. The key mediator variable in this transformation is the Rényi Differential Privacy notion, which is carefully used to harmonize privacy budgets. We also propose an innovative method to theoretically analyze the convergence for DP-FL (including our UDP-FL ) based on mode connectivity analysis. Moreover, we evaluate our UDP-FL through extensive experiments benchmarked against state-of-the-art (SOTA) methods, demonstrating superior performance on both privacy guarantees and model performance. Notably, UDP-FL exhibits substantial resilience against different inference attacks, indicating a significant advance in safeguarding sensitive data in federated learning environments.

cs.LG↗

arXiv2024

DPI: Ensuring Strict Differential Privacy for Infinite Data Streaming

Streaming data, crucial for applications like crowdsourcing analytics, behavior studies, and real-time monitoring, faces significant privacy risks due to the large and diverse data linked to individuals. In particular, recent efforts to release data streams, using the rigorous privacy notion of differential privacy (DP), have encountered issues with unbounded privacy leakage. This challenge limits their applicability to only a finite number of time slots (''finite data stream'') or relaxation to protecting the events (''event or $w$-event DP'') rather than all the records of users. A persistent challenge is managing the sensitivity of outputs to inputs in situations where users contribute many activities and data distributions evolve over time. In this paper, we present a novel technique for Differentially Private data streaming over Infinite disclosure (DPI) that effectively bounds the total privacy leakage of each user in infinite data streams while enabling accurate data collection and analysis. Furthermore, we also maximize the accuracy of DPI via a novel boosting mechanism. Finally, extensive experiments across various streaming applications and real datasets (e.g., COVID-19, Network Traffic, and USDA Production), show that DPI maintains high utility for infinite data streams in diverse settings. Code for DPI is available at https://github.com/ShuyaFeng/DPI.

cs.CR↗

arXiv2023

OptimShare: A Unified Framework for Privacy Preserving Data Sharing -- Towards the Practical Utility of Data with Privacy

Tabular data sharing serves as a common method for data exchange. However, sharing sensitive information without adequate privacy protection can compromise individual privacy. Thus, ensuring privacy-preserving data sharing is crucial. Differential privacy (DP) is regarded as the gold standard in data privacy. Despite this, current DP methods tend to generate privacy-preserving tabular datasets that often suffer from limited practical utility due to heavy perturbation and disregard for the tables' utility dynamics. Besides, there has not been much research on selective attribute release, particularly in the context of controlled partially perturbed data sharing. This has significant implications for scenarios such as cross-agency data sharing in real-world situations. We introduce OptimShare: a utility-focused, multi-criteria solution designed to perturb input datasets selectively optimized for specific real-world applications. OptimShare combines the principles of differential privacy, fuzzy logic, and probability theory to establish an integrated tool for privacy-preserving data sharing. Empirical assessments confirm that OptimShare successfully strikes a balance between better data utility and robust privacy, effectively serving various real-world problem scenarios.

cs.CR↗

arXiv2023

Efficient Privacy-Preserved Processing of Multimodal Data for Vehicular Traffic Analysis

We estimate vehicular traffic states from multimodal data collected by single-loop detectors while preserving the privacy of the individual vehicles contributing to the data. To this end, we propose a novel hybrid differential privacy (DP) approach that utilizes minimal randomization to preserve privacy by taking advantage of the relevant traffic state dynamics and the concept of DP sensitivity. Through theoretical analysis and experiments with real-world data, we show that the proposed approach significantly outperforms the related baseline non-private and private approaches in terms of accuracy and privacy preservation.

cs.CR↗

arXiv2022

Randomized Privacy Budget Differential Privacy

While pursuing better utility by discovering knowledge from the data, individual's privacy may be compromised during an analysis. To that end, differential privacy has been widely recognized as the state-of-the-art privacy notion. By requiring the presence of any individual's data in the input to only marginally affect the distribution over the output, differential privacy provides strong protection against adversaries in possession of arbitrary background. However, the privacy constraints (e.g., the degree of randomization) imposed by differential privacy may render the released data less useful for analysis, the fundamental trade-off between privacy and utility (i.e., analysis accuracy) has attracted significant attention in various settings. In this report we present DP mechanisms with randomized parameters, i.e., randomized privacy budget, and formally analyze its privacy and utility and demonstrate that randomizing privacy budget in DP mechanisms will boost the accuracy in a humongous scale.

cs.CR↗

arXiv2022

DPOAD: Differentially Private Outsourcing of Anomaly Detection through Iterative Sensitivity Learning

Outsourcing anomaly detection to third-parties can allow data owners to overcome resource constraints (e.g., in lightweight IoT devices), facilitate collaborative analysis (e.g., under distributed or multi-party scenarios), and benefit from lower costs and specialized expertise (e.g., of Managed Security Service Providers). Despite such benefits, a data owner may feel reluctant to outsource anomaly detection without sufficient privacy protection. To that end, most existing privacy solutions would face a novel challenge, i.e., preserving privacy usually requires the difference between data entries to be eliminated or reduced, whereas anomaly detection critically depends on that difference. Such a conflict is recently resolved under a local analysis setting with trusted analysts (where no outsourcing is involved) through moving the focus of differential privacy (DP) guarantee from "all" to only "benign" entries. In this paper, we observe that such an approach is not directly applicable to the outsourcing setting, because data owners do not know which entries are "benign" prior to outsourcing, and hence cannot selectively apply DP on data entries. Therefore, we propose a novel iterative solution for the data owner to gradually "disentangle" the anomalous entries from the benign ones such that the third-party analyst can produce accurate anomaly results with sufficient DP guarantee. We design and implement our Differentially Private Outsourcing of Anomaly Detection (DPOAD) framework, and demonstrate its benefits over baseline Laplace and PainFree mechanisms through experiments with real data from different application domains.

cs.CR↗

arXiv2020

R$^2$DP: A Universal and Automated Approach to Optimizing the Randomization Mechanisms of Differential Privacy for Utility Metrics with No Known Optimal Distributions

Differential privacy (DP) has emerged as a de facto standard privacy notion for a wide range of applications. Since the meaning of data utility in different applications may vastly differ, a key challenge is to find the optimal randomization mechanism, i.e., the distribution and its parameters, for a given utility metric. Existing works have identified the optimal distributions in some special cases, while leaving all other utility metrics (e.g., usefulness and graph distance) as open problems. Since existing works mostly rely on manual analysis to examine the search space of all distributions, it would be an expensive process to repeat such efforts for each utility metric. To address such deficiency, we propose a novel approach that can automatically optimize different utility metrics found in diverse applications under a common framework. Our key idea that, by regarding the variance of the injected noise itself as a random variable, a two-fold distribution may approximately cover the search space of all distributions. Therefore, we can automatically find distributions in this search space to optimize different utility metrics in a similar manner, simply by optimizing the parameters of the two-fold distribution. Specifically, we define a universal framework, namely, randomizing the randomization mechanism of differential privacy (R$^2$DP), and we formally analyze its privacy and utility. Our experiments show that R$^2$DP can provide better results than the baseline distribution (Laplace) for several utility metrics with no known optimal distributions, whereas our results asymptotically approach to the optimality for utility metrics having known optimal distributions. As a side benefit, the added degree of freedom introduced by the two-fold distribution allows R$^2$DP to accommodate the preferences of both data owners and recipients.

cs.CR↗

arXiv2018

Preserving Both Privacy and Utility in Network Trace Anonymization

As network security monitoring grows more sophisticated, there is an increasing need for outsourcing such tasks to third-party analysts. However, organizations are usually reluctant to share their network traces due to privacy concerns over sensitive information, e.g., network and system configuration, which may potentially be exploited for attacks. In cases where data owners are convinced to share their network traces, the data are typically subjected to certain anonymization techniques, e.g., CryptoPAn, which replaces real IP addresses with prefix-preserving pseudonyms. However, most such techniques either are vulnerable to adversaries with prior knowledge about some network flows in the traces, or require heavy data sanitization or perturbation, both of which may result in a significant loss of data utility. In this paper, we aim to preserve both privacy and utility through shifting the trade-off from between privacy and utility to between privacy and computational cost. The key idea is for the analysts to generate and analyze multiple anonymized views of the original network traces; those views are designed to be sufficiently indistinguishable even to adversaries armed with prior knowledge, which preserves the privacy, whereas one of the views will yield true analysis results privately retrieved by the data owner, which preserves the utility. We present the general approach and instantiate it based on CryptoPAn. We formally analyze the privacy of our solution and experimentally evaluate it using real network traces provided by a major ISP. The results show that our approach can significantly reduce the level of information leakage (e.g., less than 1\% of the information leaked by CryptoPAn) with comparable utility.

cs.CR↗