Rotated Robustness: A Training-Free Defense against Bit-Flip Attacks on Large Language Models

Dimensions overall score 8.0

• Hardware faults, specifically bit-flips in quantized weights, pose a severe reliability threat to Large Language Models (LLMs), often triggering catastrophic model collapses.technical 90%
  Evidencepartial

  Hardware faults, specifically bit-flips in quantized weights, pose a severe reliability threat to Large Language Models (LLMs), often triggering catastrophic model collapses.

  Implicationpartial

  Directly stated in the abstract as the core problem being addressed

  Verificationpartial

  partial

• The vulnerability to bit-flip attacks fundamentally stems from the spatial alignment between sensitive weight bits and extreme activation outliers.technical 85%
  Evidencepartial

  We demonstrate that this vulnerability fundamentally stems from the spatial alignment between sensitive weight bits and extreme activation outliers

  Implicationpartial

  Directly stated as the root cause explanation in the abstract

  Verificationpartial

  partial

• Rotated Robustness (RoR) reduces the stochastic collapse rate from 3.15% to 0.00% on Qwen2.5-7B under random bit-flip attacks.result 95%
  Evidencepartial

  Under random bit-flip attacks, RoR reduces the stochastic collapse rate from 3.15% to 0.00% on Qwen2.5-7B.

  Implicationpartial

  Specific numeric result directly stated in the abstract

  Verificationpartial

  partial

• RoR sustains robust reasoning on Llama-2-7B under severe targeted attacks with 50 Progressive Bit Search flips, maintaining 43.9% MMLU accuracy that nearly matches its 45.2% unattacked accuracy.result 90%
  Evidencepartial

  under severe targeted attacks with 50 Progressive Bit Search flips, RoR sustains robust reasoning on Llama-2-7B, maintaining a 43.9% MMLU accuracy that nearly matches its 45.2% unattacked accuracy

  Implicationpartial

  Specific numeric results directly stated in the abstract with clear comparison

  Verificationpartial

  partial

• Against the Single-Point Fault Attack (SPFA), RoR exponentially inflates the attack complexity from a few bits to over 17,000 precise bit-flips.result 90%
  Evidencepartial

  against the Single-Point Fault Attack (SPFA) -- the most aggressive targeted threat -- RoR exponentially inflates the attack complexity from a few bits to over 17,000 precise bit-flips.

  Implicationpartial

  Specific numeric result directly stated in the abstract

  Verificationpartial

  partial

• RoR has a negligible storage overhead of 0.31% and a minimal inference latency increase of 9.1% on Llama-2-7B.technical 95%
  Evidencepartial

  With a negligible storage overhead of 0.31% and a minimal inference latency increase of 9.1% on Llama-2-7B

  Implicationpartial

  Specific numeric results directly stated in the abstract

  Verificationpartial

  partial

• RoR is a training-free defense utilizing orthogonal Householder transformations that applies an orthogonal rotation to the activation space.method 90%
  Evidencepartial

  we propose Rotated Robustness (RoR), a training-free defense utilizing orthogonal Householder transformations. By applying an orthogonal rotation to the activation space

  Implicationpartial

  Direct description of the method's core mechanism in the abstract

  Verificationpartial

  partial

• RoR mathematically guarantees original model accuracy while providing defense against bit-flip attacks.technical 85%
  Evidencepartial

  mathematically guaranteeing original model accuracy

  Implicationpartial

  Directly stated in the abstract as a key property of the method

  Verificationpartial

  partial

Competitive landscape

Rotated Robustness offers a training-free defense against bit-flip attacks on Large Language Models, ensuring reliability and accuracy.