arXiv subjects

Liubov Nedoshivina

Publications and source records attributed to Liubov Nedoshivina.

3 recordsLinked to original sources

arXiv2026

PrivFusion: A Privacy-preserving Multi-Agent Framework for Harmonizing Distributed Datasets

The growing availability of clinical data has increased the use of machine learning, yet centralized data aggregation is often infeasible for sensitive health information. Federated Learning (FL) offers a distributed alternative, but its adoption is limited by substantial heterogeneity across institutional datasets, making harmonization a critical but frequently overlooked prerequisite for multi-site analytics. We introduce PrivFusion, a privacy-preserving multi-agent framework that automates the harmonization of structured datasets prior to federated training. PrivFusion uses agents to analyze local data, cluster semantically similar features across sites, and provide iterative transformation recommendations until alignment is achieved. Evaluation across four heterogeneous COVID-19 datasets demonstrates that PrivFusion effectively and efficiently harmonizes multi-site data while substantially reducing manual effort.

cs.LG↗

arXiv2025

Building a Foundational Guardrail for General Agentic Systems via Synthetic Data

While LLM agents can plan multi-step tasks, intervening at the planning stage-before any action is executed-is often the safest way to prevent harm, since certain risks can lead to severe consequences once carried out. However, existing guardrails mostly operate post-execution, which is difficult to scale and leaves little room for controllable supervision at the plan level. To address this challenge, we highlight three critical gaps in current research: data gap, model gap, and evaluation gap. To close the data gap, we introduce AuraGen, a controllable engine that (i) synthesizes benign trajectories, (ii) injects category-labeled risks with calibrated difficulty, and (iii) filters outputs via an automated reward model, producing large and reliable corpora for pre-execution safety. To close the guardian model gap, we propose a foundational guardrail Safiron, combining a cross-planner adapter with a compact guardian model. The adapter unifies different input formats, while Safiron flags risky cases, assigns risk types, and generates rationales; trained in two stages with a broadly explored data recipe, Safiron achieves robust transfer across settings. To close the evaluation gap, we release Pre-Exec Bench, a realistic benchmark covering diverse tools and branching trajectories, which measures detection, fine-grained categorization, explanation, and cross-planner generalization in human-verified scenarios. Extensive experiments demonstrate consistent gains of the proposed guardrail over strong baselines on Pre-Exec Bench, and ablations further distill actionable practices, providing a practical template for safer agentic systems.

cs.LG↗

arXiv2024

INTELLECT: Adapting Cyber Threat Detection to Heterogeneous Computing Environments

The widespread adoption of cloud computing, edge, and IoT has increased the attack surface for cyber threats. This is due to the large-scale deployment of often unsecured, heterogeneous devices with varying hardware and software configurations. The diversity of these devices attracts a wide array of potential attack methods, making it challenging for individual organizations to have comprehensive knowledge of all possible threats. In this context, powerful anomaly detection models can be developed by combining data from different parties using Federated Learning. FL enables the collaborative development of ML-based IDSs without requiring the parties to disclose sensitive training data, such as network traffic or sensor readings. However, deploying the resulting models can be challenging, as they may require more computational resources than those available on target devices with limited capacity or already allocated for other operations. Training device-specific models is not feasible for an organization because a significant portion of the training data is private to other participants in the FL process. To address these challenges, this paper introduces INTELLECT, a novel solution that integrates feature selection, model pruning, and fine-tuning techniques into a cohesive pipeline for the dynamic adaptation of pre-trained ML models and configurations for IDSs. Through empirical evaluation, we analyze the benefits of INTELLECT's approach in tailoring ML models to the specific resource constraints of an organization's devices and measure variations in traffic classification accuracy resulting from feature selection, pruning, and fine-tuning operations. Additionally, we demonstrate the advantages of incorporating knowledge distillation techniques while fine-tuning, enabling the ML model to consistently adapt to local network patterns while preserving historical knowledge.

cs.CR↗