arxiv arXiv cs.LG · 90d ago · research

Zero-Overhead Telemetry Detects Hidden ML Training

from English

A study evaluates GPU workload classification using only zero-overhead NVML telemetry. The classifier achieves 98.2% accuracy in identifying training workloads and 43-87% accuracy against adversarially disguised, unexpected workloads across 9 GPU models.

Importance 3/3 New harness with differentiators arXiv cs.LG NVIDIA Evaluation & benchmarks Hardware & chips Safety & alignment

Read original

lab NVIDIA Technical Blog · 72d ago

NVIDIA introduces Confidential Computing for secure AI inference

NVIDIA has engineered NVIDIA Confidential Computing (CC) as a solution to address data privacy, sovereignty, and security concerns during AI model inference. The technology is designed to provide a secure environment for processing data while it is in use, specifically targeting the needs of agentic AI applications. This approach aims to enable organizations to adopt AI without compromising on security or performance.

arxiv arXiv cs.AI · 89d ago · 6 views

CRAX: Fast Safe Reinforcement Learning Benchmarking

CRAX introduces a high-fidelity, accelerated safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.

arxiv arXiv cs.LG · 89d ago · 5 views

CRAX: Fast Safe Reinforcement Learning Benchmarking

CRAX introduces a high-fidelity, fast safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.

lab NVIDIA Research · 11h ago · 3 views

ShareMMU enables secure address translation sharing among untrusted accelerators with low overhead

Researchers present ShareMMU, an IOMMU design that allows multiple untrusted accelerators to securely reuse pre-translated addresses within a shared virtual address space. This approach mitigates side-channel risks associated with large Shared Translation Lookaside Buffers (TLBs) while maintaining high performance.

media Hugging Face Forums · 1d ago · 4 views

Qwen-Scope and Gemma Scope 2 show evaluation awareness is eight directions, not one

A study using EvalAwareBench tests whether language models use a single shared direction or multiple cue-specific detectors to identify evaluation contexts. The research toggled eight trigger factors independently across 1,298 prompts while holding underlying tasks and entities fixed.