LATIDIA · Ciberseguridad

Robustez rotada: una defensa sin entrenamiento contra ataques de bit-flip en modelos de lenguaje grandes

arXiv: 2603.16382v2Tipo de Anuncio: reemplazar Resumen: La corrupción de bit-flip de pesos cuantificados representa una seria amenaza de confiabilidad para los Modelos de Lenguaje Grande (LLM), ya que incluso un pequeño número de fallas de peso puede desencadenar catastr

La noticia

arXiv:2603.16382v2 Announce Type: replace Abstract: Bit-flip corruption of quantized weights poses a serious reliability threat to Large Language Models (LLMs), as even a small number of weight faults can trigger catastrophic model degradation. We show that such failures can be strongly amplified when corrupted weights are coupled to high-sensitivity activation channels. Based on this observation, we propose Rotated Robustness (RoR), a training-free sensitivity-aware hierarchical rotation that applies matched orthogonal transformations to activations and weights, redistributing sensitive channel contributions across feature dimensions while preserving the underlying linear mapping in exact arithmetic. Across eight evaluated LLMs, RoR matches or exceeds all compared methods in the number of cumulative PBS flips sustained before