Federated Learning Tackles Data Heterogeneity in Malware Image Classification
A new arXiv paper examines how federated learning can be applied to malware image classification when training data is unevenly distributed across clients.
Research Digest
Preprints and papers, rewritten in plain English.
LLMs Misjudge CVSS Severity When Bug Reports Include Narratives
A new empirical study finds that the narrative text in bug bounty reports can lead large language models to assign inaccurate CVSS severity scores.
Migrating IPsec to Post-Quantum Crypto: Composite Key Sharing in IKEv2
A new arXiv paper proposes reliable composite key sharing mechanisms in IKEv2 to ease IPsec's migration to post-quantum cryptography.
Formalizing Rényi Differential Privacy in Isabelle/HOL
A new draft paper uses the Isabelle/HOL proof assistant to give machine-checked formalization of Rényi differential privacy, a variant that provides tight privacy bounds for machine learning.
PRICE Framework Uses Homomorphic Encryption for Private Image Classification
A new framework called PRICE aims to let mobile apps classify images in the cloud without exposing raw image data.
RAGLeak: Supply-Chain Backdoor Targets RAG Knowledge Bases
A new paper describes a privacy backdoor that exploits retrieval-augmented generation's reliance on external knowledge bases.
SpanGuard proposes lighter indirect prompt injection defense
A new approach separates finding instructions from deciding whether to obey them, aiming to make guardrails work on smaller models.
Spotting Security Hallucinations in AI-Generated Code
A new technique called SPLASH targets code that passes functional tests but contains hidden security flaws.
NIOCIS: A Routing Layer to Cut Repetitive Thinking in Tool-Use Agents
A new routing layer called NIOCIS aims to reduce the cost of tool-use agents by folding execution logs into decision-making, avoiding repeated re-reading and thinking at each hop.
PhantomTac: Spoofing Robotic Touch with Deploy-and-forget Tags
A new attack shows how magnetic tactile sensors in robots can be fooled by small tags that create false touch signals.
Fake News Can Crash LLM Trading Agents
A new study shows that LLM-based trading agents are vulnerable to information injection attacks, where a single fabricated headline can trigger real market crashes.
LeakBench: Auditing Reward Hacking in Clinical Modeling Agents
A new executable audit suite, LeakBench, systematically tests whether clinical AI agents cheat by exploiting prediction-time leakage.
Fortress Uses Social Graphs to Make Transparency Logs Fork-Resistant
A new system called Fortress leverages social graph structure to prevent undetected forks in transparency logs.
GreenNIDS Benchmark Tracks Performance and Cost of NetFlow Intrusion Detection
A new workflow and dataset lets intrusion-detection researchers compare predictive accuracy alongside computational cost.
Exact Formula for Multivariate Moments Under Differential Privacy
A new arXiv preprint shows how randomized response distorts multivariate cross-moments, with an exact attenuation formula that could make privatized statistics easier to correct.
Cost-aware metric proposed for evaluating iterative LLM jailbreaks
A new evaluation approach accounts for the repeated calls that real-world jailbreak attempts require, going beyond simple attack success rate.
Formal Treatment of Repaired Dual-Modulus RSA Countermeasure
A new analysis revisits the broken DM-RSA scheme and formally treats its repaired variant, DM-RSA-OAEP, as a fault-detection countermeasure.
Shielding Antivirus Signatures from Black-Box Knowledge Extraction
A proposed method aims to keep antivirus detection knowledge safe from reverse-engineering via repeated queries.