[论文解读] Position: Certifiable State Integrity in Cyber-Physical Systems -- Why Modular Sovereignty Solves the Plasticity-Stability Paradox

Enzo Nicolás Spotorno, Antônio Augusto Medeiros Fröhlich|arXiv (Cornell University)|Jan 29, 2026

Adversarial Robustness in Machine Learning被引用 0

一句话总结

本文认为在 CPS 的可认证状态完整性需要一个模块化主权方法（HYDRA），通过冻结的 regime-specific 专家库和对不确定性的融合来解决可塑性-稳定性悖论，并实现跨生命周期转换的可验证安全性。

ABSTRACT

The machine learning community has achieved remarkable success with universal foundation models for time-series and physical dynamics, largely overcoming earlier approximation barriers in smooth or slowly varying regimes through scale and specialized architectures. However, deploying these monolithic models in safety-critical Cyber-Physical Systems (CPS), governed by non-stationary lifecycle dynamics and strict reliability requirements, reveals persistent challenges. Recent evidence shows that fine-tuning time-series foundation models induces catastrophic forgetting, degrading performance on prior regimes. Standard models continue to exhibit residual spectral bias, smoothing high-frequency discontinuities characteristic of incipient faults, while their opacity hinders formal verification and traceability demanded by safety standards (e.g., ISO 26262, IEC 61508). This position paper argues that the plasticity-stability paradox cannot be fully resolved by global parameter updates (whether via offline fine-tuning or online adaptation). Instead, we advocate a Modular Sovereignty paradigm: a library of compact, frozen regime-specific specialists combined via uncertainty-aware blending, which we term "HYDRA" (Hierarchical uncertaintY-aware Dynamics for Rapidly-Adapting systems). This paradigm ensures regime-conditional validity, rigorous disentanglement of aleatoric and epistemic uncertainties, and modular auditability, offering a certifiable path for robust state integrity across the CPS lifecycle.

研究动机与目标

• Identify theoretical and practical barriers to deploying universal foundation models in safety-critical CPS.
• Propose Modular Sovereignty (HYDRA) as a paradigm to decouple library constitution from runtime arbitration and resolve forgetting.
• Outline architectural principles, advantages, and open challenges toward lifecycle-certifiable physical learning.
• Enable regime-conditional validity, disentangled uncertainties, and modular auditability for safety-standard certification.

提出的方法

• Define State Integrity as continuous, physically interpretable linkage between physical and digital representations across lifecycle.
• Introduce HYDRA: a library of frozen, regime-specific specialists (S_k) combined by an uncertainty-aware Governor (B) and guarded by an Integrity Monitor (I).
• Formalize output as a convex combination: ŷ_t = sum_k π_k^{(t)} S_k(x_t) with π in the simplex, ensuring the state remains in the convex hull of valid local manifolds.
• Phase I: Offline constitution of L with Type-I (Physics-Derived) and Type-II (Data-Derived) specialists and a Constitutional Greedy Accretion vetting process.
• Phase II: Online calibration where the Governor infers low-dimensional mixing coefficients π_t via residuals and uncertainty metrics, enabling regime-conditional conformal prediction.
• Tie-ins to LPV theory via polytopic generalization and RPI Zonotopes for safety sets; employ Dirichlet priors (α) to manage sparsity vs. continuity in regime switching.

Figure 1 : Simplex Integration. The Governor isolates the AI (QM) from the Safety Core (ASIL D). Specialists provide estimation; the Governor and fallback form the high-assurance safety channel. Figure crafted with the help of a GenAI Image Model.

实验结果

研究问题

• RQ1 How can we maintain regime-conditional validity and certifiability in non-stationary CPS without catastrophic forgetting?
• RQ2 Can a library of frozen regime-specific specialists combined by an uncertainty-aware governor provide auditable safety guarantees under ISO/IEC safety standards?
• RQ3 How does residual-based arbitration enable reliable detection of regime shifts, faults, and aging in safety-critical dynamics?
• RQ4 What architectural and mathematical tools (e.g., LPV, polytopic theory, zonotopes) support modular certification and tractable verification?

主要发现

• Monolithic foundation models exhibit plasticity-stability and certifiability challenges in non-stationary CPS, including forgetting and opaque verification.
• HYDRA decouples library constitution from runtime arbitration, enabling fault-tolerant, certifiable adaptation via a small set of frozen specialists.
• Uncertainty-aware blending (Governor) uses runtime residuals to detect regime shifts and manage switching with controllable ambiguity, leveraging LPV-like guarantees.
• Convex combination of specialists preserves state within the convex hull of valid manifolds, supporting modular verification and safety envelopes (RPI Zonotopes).
• The Integrity Triangle (Governance Health, Uncertainty, and Transparency) provides a structured runtime assurance framework with simple-safety fallbacks (Simplex Architecture).
• Phase-wise library construction and vetting minimize forgetting, enable independent certification of components, and improve availability over monolithic wrappers.

Figure 2 : The Architecture of Sovereignty. Illustrated via the Vehicle Dynamics case study. The framework separates the offline Constitution (Library of Specialists) from the online Governance (in this example, implementing a Bayesian method for uncertainty-quantification, such as variational infer