Search posts, papers, and topics
Hacker News front page1 min readintro
The article discusses techniques for exfiltrating machine learning model weights. It highlights the security implications for ML deployments.
ML engineers and security teams should care about protecting model weights from exfiltration.
2/10Decoy Direction Optimization (DDO) is a post‑hoc weight‑editing defense for open‑weight LLMs that injects a high‑magnitude nonlinear decoy into MLP neurons, corrupting contrastive estimators used by Refusal Feature Ablation (RFA) attacks. The paper proves a spectral bound on the effect, evaluates DDO on six model families (including Llama‑3‑8B‑Instruct), and shows <10 % attack success rate (ASR)…
Cloudflare’s Page Shield uses a graph‑neural‑network (GNN) to model JavaScript as a syntax‑tree graph, followed by a lightweight LLM for second‑opinion triage and an ensemble of frontier models for deep analysis. This pipeline caught eight malicious payloads across four distinct affiliate‑theft and backdoor techniques that traditional scanners missed, demonstrating the need for runtime, behavior‑…
Yelp added an Inference Plugin to Nrtsearch that runs XGBoost and neural‑network models inside the search engine, eliminating a separate scoring service. The plugin extracts features from index documents, loads MLeap bundles from MLflow, and serves predictions on replica nodes with millisecond latency.
Cloudflare adds a Disallow AI Training setting that lets sites stay indexed for search while blocking crawlers from using their content for AI model training. The feature replaces the generic Block AI Bots option with granular Search, Training, and Agent controls.
The paper presents Infinite-Parameter LLMs, where a compact hypernetwork creates feed‑forward weights from live user data and updates a Bayesian latent code online, keeping the stored model size constant while effectively having infinite parameters. This design aims to improve over standard in‑context learning and retrieval by persisting knowledge in weights and freeing context space.
Dynamically Scaled Activation Steering (DSAS) is a method‑agnostic framework that learns per‑token, per‑layer scaling factors to turn existing activation‑steering interventions on only when a model is likely to produce undesired output (e.g., toxic text). The scaling can be optimized jointly with any steering function, improves the toxicity‑utility trade‑off on language models, transfers to text‑…
