2786 papers
This paper presents Agent Guard, a Linux reference monitor that enforces human-authorized damage boundaries for AI agents by using eBPF-based LSMs to deterministically deny unauthorized file and network access while propagating strict "no-egress" states across process hierarchies, achieving significantly lower overhead than existing baselines.
This paper introduces Mastyf Guard 1.5B, a lightweight, capability-based security framework that implements a "Cognitive Harvard Architecture" to physically decouple data ingestion from action execution, thereby mathematically guaranteeing immunity to Indirect Prompt Injection while achieving 99.33% threat recall and sub-millisecond latency on standard CPU hardware.
This paper introduces SmartADAPT-Net, a lightweight neural framework that enhances smartwatch-based activity recognition in free-living environments by combining a fixed population model with a two-stage adaptation strategy: an immediate observation-conditioned correction and a progressive personalization based on user-confirmed evidence, achieving high accuracy without requiring deployment-time retraining.
This paper proposes a novel hybrid ransomware detection framework that integrates multi-layer behavioral analysis with a stacking ensemble of classical machine learning models and LSTM networks, achieving 99% accuracy and superior generalization by effectively capturing both discriminative patterns and temporal dependencies in attack sequences.
This paper synthesizes findings from 41 empirical studies to propose a comprehensive, multi-layered validation framework that progresses from data leakage prevention and interpretability audits to conformal calibration and execution-grounded agentic testing, addressing critical failure modes like class imbalance and unfaithful attribution that standard validation methods overlook in high-stakes applications.
This study evaluates various deep learning models for skin cancer classification using the HAM10000 dataset, demonstrating that a hybrid model and ResNet50 achieve superior accuracy (99.78% and 99.88%, respectively) compared to other techniques, thereby highlighting their potential to enhance early diagnosis and resource allocation in dermatology.
This paper proposes the T3-DEF framework, a socio-technical analytical model that identifies a critical inverse relationship between rising automation density and declining verification diversity in AI-assisted systems, revealing a threshold near 0.58 where recursive operational vulnerability accelerates and emphasizing the necessity of maintaining distributed oversight structures to ensure safety.
This paper proposes an Axiomatic Design-based, open-source toolkit that integrates a graphical user interface and agentic AI to systematically transform fragmented, manual data preparation in smart manufacturing into a reproducible, scalable scientific discipline, thereby enabling advanced downstream applications like digital twins.
The paper proposes Multi-View Dual-Contrastive Graph Learning (MDCGA), a lightweight framework that integrates lexical, semantic, and keyword-based relational signals through enhanced dual contrastive objectives to achieve robust semi-supervised text classification with few labeled examples by mitigating false negatives and unreliable augmentations.
This paper proposes an AI-driven, metaheuristic-optimized access selection framework for secure MPLS-based heterogeneous networks that utilizes a lightweight MLP to predict video QoE and a fitness-based SDN controller to dynamically select optimal LTE/Wi-Fi handover targets, thereby significantly improving video continuity, network performance, and energy efficiency compared to conventional RSSI-based strategies.