SMART framework measures semantic impact of speech-recognition errors
A new evaluation framework for speech recognition goes beyond word error rate to label which transcription mistakes actually change meaning.
Research Digest
Preprints and papers, rewritten in plain English.
WorldSR grounds image super-resolution with world knowledge search
A new arXiv preprint proposes countering the stochasticity of generative priors by grounding super-resolution in explicitly retrieved world knowledge.
MAS-Guard Proposes Unified Safety Guardrails for Multi-Agent LLMs
A new arXiv preprint outlines a unified guardrail framework to mitigate safety risks in agentic AI systems that plan and execute complex workflows.
Secure Delegation Networks Could Fix Liquid Democracy's Power Imbalances
A new arXiv paper proposes secure delegation networks to mitigate voting power concentration in liquid democracy systems.
WaveCommitments: Private Trading for Speech Agents
A new arXiv paper proposes a commitment scheme to let autonomous speech agents trade data without revealing it.
nokast-secureRAG: Small LM-Based Defense for RAG Prompt Injection
A new arXiv preprint proposes using small language models to detect indirect prompt injection and retrieval poisoning in local RAG systems.
Role-Aware Disclosure for LLMs in Mixed-Sensitivity Contexts
A new arXiv paper proposes a role-aware, graduated approach to help LLMs handle mixed-sensitivity information without over-disclosing.
Polynomial separation found between pure and entangled pebbling
The result could affect security arguments for memory-hard functions and proofs of space based on pebbling lower bounds.
Digital Twin Approach Aims to Address ICS Cyber Threats
A new arXiv paper proposes using digital twins to improve cyber-social impact as ICS and IT integration expands the threat landscape.
Interflow Targets Cost Barrier for Private Edge AI
A new approach aims to reduce the overhead that has kept privacy-preserving machine learning from practical edge deployment.
NetGuise Paper Targets Early-Stage Traffic Classifiers
A new arXiv preprint proposes an adversarial traffic generation method that could evade classifiers making decisions from only the first few packets.
Privacy Accounting Method Targets Empirical Denoisers in Diffusion Models
A new approach to privacy accounting for empirical denoisers could help diffusion models achieve formal guarantees without the full cost of DP training.
Predicting When More Fuzzing Is Necessary in Continuous Fuzzing
A new arXiv paper tackles the challenge of deciding whether a continuous fuzzing campaign needs additional runs to catch more vulnerabilities.
Execution-Time Validation for Stale Commands in Industrial CPS
New research argues that industrial control commands must be re-validated at execution time because the plant may have changed while the command was in transit.
Analyzing Cross-Language Data Flows in Android Hybrid Apps
A new framework targets information leaks that arise when Android native code and WebView JavaScript exchange data.
Shimura Class Group Actions on Abelian Varieties for Post-Quantum Crypto
A new arXiv paper explores how Shimura class group actions on oriented superspecial abelian varieties could serve as a basis for post-quantum cryptography.
Watermarking Could Thwart LLM Theft by Surrogate Models
A new arXiv proposal embeds robust watermarks in proprietary LLM outputs to deter training surrogate copies with minimal semantic change.
Remote Activation of LiDAR Trojan Malware via Modulated Signals
A new arXiv study reveals a previously unexplored attack surface that could let attackers remotely trigger dormant malware in LiDAR sensing pipelines.