SoK: Resilience and Fault Tolerance in Cyber-Physical Systems

Rahul Bulusu · 2025

Read SoK: Resilience and Fault Tolerance in Cyber-Physical Systems by Rahul Bulusu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-20873.txt.

Introduction

Rahul Bulusu surveys nearly twenty years of research on keeping cyber-physical systems running through faults. Such systems, from power grids to medical devices and robots, fuse software tightly with physical motion. A small glitch in a sensor or network can cascade into physical instability in ways single-layer thinking misses.

Bulusu organizes the field with an Origin-Layer-Effect taxonomy: where faults start, which layer they cross, and what they finally do. Mapping well-known systems onto it exposes shared failure pathways and blind spots. Four structural gaps recur, including learning-based control lacking stability guarantees and weak forensic visibility. The survey closes with directions joining robust control, monitoring, and formal assurance.

Catalog record

CREATORRahul Bulusu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-20873

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SoK: Resilience and Fault Tolerance in Cyber-Physical Systems Rahul Bulusu Abstract Cyber-Physical Systems (CPS) now support critical infrastructure spanning transportation, energy, manufacturing, medical devices, and autonomous robotics. Their defining characteristic is the tight coupling between digital computation and continuous physical dynamics which enables sophisticated autonomy but also creates highly non-linear failure modes. Small disturbances at sensors, firmware, networks, or physical interfaces can propagate through estimation and control pipelines, producing cascading instabilities that defy traditional single-layer reasoning. This Systematization of Knowledge (SoK) unifies nearly two decades of CPS resilience research into a structured Origin–Layer–Effect (OLE) taxonomy. This taxonomy provides a cross-layer lens for understanding how faults arise, how they propagate, and why unrelated CPS failures often share deep structural similarities. By mapping representative systems including RockDrone, MAYDAY, M2MON, HACMS, Byzantine fault-tolerant control, and learning-based recovery mechanisms onto the taxonomy, we reveal patterns of coverage, persistent blind spots, and recurring pathways of fault amplification. Our analysis identifies four structural gaps that span multiple CPS domains: (1) physical-model manipulation, (2) ML-enabled control without stability guarantees, (3) semantic inconsistencies between formal models and firmware, and (4) inadequate forensic visibility across cyber and physical layers. These insights motivate new directions for resilient CPS design, integrating robust control, runtime monitoring, formal assurance, and system-level visibility. 1 Introduction Cyber-Physical Systems (CPS) now exists in our daily lives, governing domains in which digital decisions directly influence the physical world. Unlike conventional software, CPS mishaps lead not merely to data loss but to real, tangible harm. Drones crash onto crowds, autonomous vehicles misinterpret road conditions, industrial machinery fails catastrophically, and medical devices deliver incorrect dosages. These failures arise because CPS tightly integrate sensing, computation, actuation, and physical dynamics, forming complex feedback loops that amplify small perturbations into full-system failures. Errors that remain inconsequential in traditional computing environments can trigger catastrophic outcomes when they interact with the dynamics of the physical plant. To illustrate these risks, consider the RockDrone attack [ 21 ] , where a single acoustic tone excites mechanical resonances in a MEMS gyroscope, producing biased angular velocity estimates. These small sensor perturbations propagate through the estimator and into the attitude controller, ultimately destabilizing the UAV. This illustrates the central challenge of CPS resilience: faults originating in one layer can cascade nonlinearly through the entire feedback loop. The expansion of autonomy has intensified these challenges. Modern CPS depend on heterogeneous sensing modalities, distributed embedded systems, wireless communication networks, cloud-assisted computation, and learning-enabled components for perception and planning. Each of these layers introduces its own assumptions about timing, trust, noise, and error bounds. When these assumptions are violated by environmental disturbances, hardware degradation, or adversarial manipulation, the resulting behavior often escapes the abstractions of any single discipline. Resilience, in this context, means more than traditional robustness or reliability. A resilient CPS must maintain acceptable behavior under both stochastic disturbances and strategically adversarial actions. It must also provide operators with sufficient visibility to understand why failures occur so that future systems can avoid repeating them. Despite substantial progress across multiple communities like robust control, sensor fusion, embedded systems, security, and formal verification, CPS resilience research remains fragmented. 1.1 The Need for Systematization The research community has produced a rich but fragmented body of work on CPS resilience. Studies examine GPS spoofing [ 15 ] , LiDAR illusions [ 19 ] , acoustic gyroscope manipulation [ 21 ] , CAN bus injection [ 6 ] , timing interference [ 24 ] , firmware vulnerabilities, and formal verification of UAV platforms [ 9 ] . Each provides valuable depth on one part of the resilience problem, yet CPS failures are inherently cross-layer phenomena. This motivates a Systematization of Knowledge (SoK) that unifies these disparate threads using a structural Origin–Layer–Effect taxonomy. Rather than enumerating attacks or defenses, we map them into this design space, exposing structural patterns and blind spots that are not visible when examining papers individually. 1.2 Contributions This SoK makes three primary contributions. First, we propose a cross-layer taxonomy for CPS resilience that explicitly factors (1) the origin of a disturbance, (2) the architectural layer where it manifests, and (3) the effect type . This provides a common vocabulary spanning control theory, embedded systems, and security. Second, we map representative CPS resilience frameworks including MAYDAY [ 25 ] , M2MON [ 17 ] , RockDrone [ 21 ] , Learn2Recover [ 5 ] , HACMS [ 9 ] , and Byzantine fault-tolerant control [ 12 ] onto this taxonomy. This reveals overlaps, conflicts, and unprotected design regions. Third, we analyze UAVs, autonomous vehicles, industrial systems, and medical CPS to extract cross-domain structural patterns. Across domains, CPS failures often stem from shared structural weaknesses such as sensor trust assumptions, timing sensitivities, model–firmware mismatches, and insufficient forensic visibility. 1.3 Scope and Methodology We draw from control, embedded systems, CPS security, robotics, and industrial reports. We include work that (i) involves closed-loop CPS dynamics, (ii) captures multi-layer fault propagation, or (iii) proposes defenses informed by physical modeling. We exclude purely IT security and purely control theory unless they materially inform CPS resilience. We recursively expanded our collection starting from canonical CPS attacks (e.g., GPS spoofing, CAN injection, acoustic sensor attacks) and CPS defense mechanisms (e.g., Simplex, runtime monitoring, robust estimation). Each paper was classified along our taxonomy axes and used to refine the taxonomy. 2 Background Cyber-Physical Systems are defined by closed-loop interactions between the physical world and computational processes. Understanding resilience requires examining the structure of this loop and the foundational principles governing its stability. Terminology. Throughout this paper, we distinguish between three related but distinct concepts. A fault is any deviation in the sensor, actuator, or internal software signals. An error is a resulting deviation in the estimated system state. A failure occurs when errors propagate to violate a safety- or mission-critical invariant. Adversarial attacks induce faults intentionally, while control-semantic faults arise when the physical and computational interpretations of system behavior diverge. 2.1 CPS Architectural Structure A CPS typically consists of five interacting subsystems: physical processes, sensors, computational controllers, actuators, and communication networks. The physical process evolves continuously according to natural laws or engineering dynamics. Sensors observe physical variables such as position, velocity, temperature, pressure, or acceleration, but measurements are corrupted by noise, bias, drift, and environmental effects [ 11 ] . Controllers attempt to estimate the system state and compute actuator commands that achieve a control objective, such as stabilizing a drone, tracking a lane, or maintaining pressure within a safe band. Actuators apply forces or changes that alter the physical world. Communication channels relay sensor readings, actuator commands, and state estimates, often under bandwidth or latency constraints and sometimes over insecure protocols like CAN or Modbus [ 6 ] . In theory, CPS stability emerges from carefully designed feedback loops in which controllers are synthesized using accurate models of plant dynamics and well-characterized disturbance profiles. In practice, real-world deployments deviate from these idealized assumptions. Noise is rarely Gaussian, dynamics is nonlinear and time-varying, communication is unreliable, and environmental conditions change over seasons, weather patterns, and operational cycles. These deviations compromise theoretical guarantees and open windows for adversarial interference. 2.2 Control-Theoretic Principles Control theory provides mathematical guarantees about the behavior of dynamic systems. Stability ensures that errors decay over time

Read the complete text