MedRoundsQA: A Persona and Difficulty Aware Evaluation for Multi-Turn Medical Consultations
针对单轮医疗咨询无法反映真实情况的问题,通过构建多轮对话基准MedRoundsQA并分类难度来评估模型性能。
针对单轮医疗咨询无法反映真实情况的问题,通过构建多轮对话基准MedRoundsQA并分类难度来评估模型性能。
This work addresses the information leakage inherent in multi-stage masked NTT pipelines under standard modular reduction in post-quantum cryptographic hardware. We propose a security quantification method independent of pipeline depth, leveraging the Lean 4 formal verification framework to model arbitrary k-stage PF-PINI(≤2) masked NTT implementations with fresh inter-stage masks. For the first time, we formally prove that Montgomery modular reduction satisfies PF-PINI(2) security. By introducing a novel k-stage composition theorem, we demonstrate that the overall security bound is determined solely by the final stage, yielding a unified upper bound on conditional leakage probability of 2/q. This result has been machine-verified, closing the analysis loop within the QANARY platform and providing rigorous, verifiable security guarantees for masked NTT hardware designs.
This work addresses the lack of a composability theory for arithmetic masking over prime fields, which undermines side-channel security guarantees for NTT-based post-quantum cryptographic hardware in multi-stage pipelines. The authors propose a Prime-Field Probing-Isolating Non-Interference (PF-PINI) model that introduces an inter-stage “mask refreshing” mechanism to fully uniformize intermediate values, thereby enabling secure composition. They establish the first machine-verified composability theorem for prime-field arithmetic masking, formally proving that mask refreshing completely eliminates the influence of prior-stage security parameters and integrates faithfully with a hardware-aware model of Barrett reduction. Using Lean 4, they complete 18 gap-free formal proofs and apply a combined structural-dependence and probing model to diagnose hardware accelerators. The methodology successfully identifies a first-order leakage in Microsoft’s Adams Bridge PQC accelerator caused by missing inter-stage refreshing, aligning with empirical observations.
This work addresses the absence of a formal characterization of information leakage in first-order masked Barrett reduction over prime fields under side-channel attacks. The authors propose and machine-verify a “1-bit barrier” property, establishing that each internal wire leaks at most one bit of information for any modulus. They formulate a trichotomy theorem on the cardinalities of preimages under internal wire mappings and prove that the maximum multiplicity does not exceed two. Furthermore, they introduce the PF-PINI security model and demonstrate that the implementation satisfies PF-PINI(2). Using Lean 4 and Mathlib, they formally verify twelve theorems, providing the first general, machine-checked bounds on mask cardinalities and leakage analysis for Barrett reduction applicable to both ML-KEM and ML-DSA.
This study investigates whether stage-wise arithmetic masking in number-theoretic transform (NTT)-based post-quantum cryptographic hardware pipelines ensures end-to-end side-channel security under r-bearing scenarios with fresh randomness. For the first time, we formally verify in Lean 4 and Mathlib the arithmetic masking scheme of Cooley–Tukey NTT butterfly operations over ℤ_q, proving that a k-stage pipeline satisfies stage-wise uniformity under the ISW first-order probing model, yet fails to guarantee pointwise value independence. Our work bridges the formal verification gap for compositional security of arithmetic masking beyond Boolean domains and exposes a structural vulnerability in the Adams Bridge accelerator stemming from its violation of the fresh-masking assumption. All nine theorems are fully machine-verified with zero “sorry” placeholders across 1,738 build tasks.
针对单轮医疗咨询无法反映真实情况的问题,通过构建多轮对话基准MedRoundsQA并分类难度来评估模型性能。
This work addresses the information leakage inherent in multi-stage masked NTT pipelines under standard modular reduction in post-quantum cryptographic hardware. We propose a security quantification method independent of pipeline depth, leveraging the Lean 4 formal verification framework to model arbitrary k-stage PF-PINI(≤2) masked NTT implementations with fresh inter-stage masks. For the first time, we formally prove that Montgomery modular reduction satisfies PF-PINI(2) security. By introducing a novel k-stage composition theorem, we demonstrate that the overall security bound is determined solely by the final stage, yielding a unified upper bound on conditional leakage probability of 2/q. This result has been machine-verified, closing the analysis loop within the QANARY platform and providing rigorous, verifiable security guarantees for masked NTT hardware designs.
This work addresses the lack of a composability theory for arithmetic masking over prime fields, which undermines side-channel security guarantees for NTT-based post-quantum cryptographic hardware in multi-stage pipelines. The authors propose a Prime-Field Probing-Isolating Non-Interference (PF-PINI) model that introduces an inter-stage “mask refreshing” mechanism to fully uniformize intermediate values, thereby enabling secure composition. They establish the first machine-verified composability theorem for prime-field arithmetic masking, formally proving that mask refreshing completely eliminates the influence of prior-stage security parameters and integrates faithfully with a hardware-aware model of Barrett reduction. Using Lean 4, they complete 18 gap-free formal proofs and apply a combined structural-dependence and probing model to diagnose hardware accelerators. The methodology successfully identifies a first-order leakage in Microsoft’s Adams Bridge PQC accelerator caused by missing inter-stage refreshing, aligning with empirical observations.
This work addresses the absence of a formal characterization of information leakage in first-order masked Barrett reduction over prime fields under side-channel attacks. The authors propose and machine-verify a “1-bit barrier” property, establishing that each internal wire leaks at most one bit of information for any modulus. They formulate a trichotomy theorem on the cardinalities of preimages under internal wire mappings and prove that the maximum multiplicity does not exceed two. Furthermore, they introduce the PF-PINI security model and demonstrate that the implementation satisfies PF-PINI(2). Using Lean 4 and Mathlib, they formally verify twelve theorems, providing the first general, machine-checked bounds on mask cardinalities and leakage analysis for Barrett reduction applicable to both ML-KEM and ML-DSA.
This study investigates whether stage-wise arithmetic masking in number-theoretic transform (NTT)-based post-quantum cryptographic hardware pipelines ensures end-to-end side-channel security under r-bearing scenarios with fresh randomness. For the first time, we formally verify in Lean 4 and Mathlib the arithmetic masking scheme of Cooley–Tukey NTT butterfly operations over ℤ_q, proving that a k-stage pipeline satisfies stage-wise uniformity under the ISW first-order probing model, yet fails to guarantee pointwise value independence. Our work bridges the formal verification gap for compositional security of arithmetic masking beyond Boolean domains and exposes a structural vulnerability in the Adams Bridge accelerator stemming from its violation of the fresh-masking assumption. All nine theorems are fully machine-verified with zero “sorry” placeholders across 1,738 build tasks.