RoS-Guard: Robust and Scalable Online Change Detection with Delay-Optimal Guarantees

Ze‐Lin Zhu, Yancheng Huang, Kai YangarXiv 2025 · DOI · arXiv

In the world of streaming data—whether it's monitoring a smart grid for cyber-attacks or a wireless network for signal blockages—the ability to spot a "change" the moment it happens is critical. This field, known as Online Change Detection (OCD), faces a persistent double-bind: real-world systems are messy and unpredictable, and as these systems grow larger, the math required to monitor them becomes agonizingly slow.

A new paper from researchers at Tongji University introduces RoS-Guard, a framework designed to break this deadlock by combining robust optimization with the parallel processing power of modern GPUs.

TL;DR

RoS-Guard is a robust and scalable algorithm for detecting abrupt changes in streaming data, specifically designed for linear systems where the underlying model is uncertain. By reformulating complex optimization problems into a "neural unrolling" architecture, it achieves a 20x computational speedup in large-scale scenarios while providing rigorous theoretical guarantees on detection delay and false alarm rates.

What Problem Does This Paper Address?

Most OCD algorithms operate on a dangerous assumption: that we have a perfect mathematical model of the system. In reality, factors like environmental fluctuations in a power grid or estimation errors in a wireless channel mean our "nominal" model is often slightly off. When an algorithm assumes a perfect model but encounters a fuzzy one, it either misses real changes or cries wolf with constant false alarms.

Furthermore, detecting changes optimally usually involves solving a Generalized Log-Likelihood Ratio (GLLR). In high-dimensional systems (like a massive sensor network), calculating this ratio in real-time becomes a computational bottleneck. Existing methods either sacrifice accuracy for speed or become too slow to be truly "online."

Methods

The authors tackle these issues through a multi-step mathematical "makeover":

  1. Robust Reformulation: Instead of assuming a fixed system matrix, they assume the system belongs to an uncertainty set (like a polyhedron or ellipsoid). They reformulate the GLLR as a Mixed-Integer Quadratic Programming (MIQP) problem to account for this uncertainty.
  2. Tight Relaxation: Since MIQP is notoriously hard to solve quickly (NP-hard), they apply a Second-Order Cone Programming (SOCP) relaxation. This simplifies the math into a convex form that is much faster to solve while remaining "tighter" (more accurate) than standard industry relaxations.
  3. Neural Unrolling: This is the secret sauce for speed. The researchers "unroll" the iterative optimization steps into layers of a neural network. This allows the algorithm to run on a GPU, performing parallel computations that would take a CPU significantly longer. Unlike a "black box" AI, this network is structured by the physics of the optimization problem itself.

Datasets

The researchers validated RoS-Guard on two distinct, high-stakes scenarios:

  • Smart Grid (Dataset I): Using the standard IEEE 14-bus system benchmark, they simulated False Data Injection Attacks (FDIAs). They modeled uncertainties in line admittances caused by environmental changes.
  • Wireless MIMO (Dataset II): They used a 2×4 MIMO testbed (based on USRP devices) to detect channel blockages caused by humans, drones, or vegetation. This represents a classic "change" in a dynamic wireless environment.

Key Insights and Ideas

  • The Orthogonal Complement Trick: The authors realized that some changes are "hidden" within the system's normal operations. By projecting observations onto the orthogonal complement of the system's column space, they can isolate the "detectable" signals from the "unobservable" ones.
  • Optimization as an Architecture: By using neural unrolling, the authors treat the optimization's step sizes as learnable parameters. This means the algorithm "learns" how to solve the specific detection problem faster over time without losing the theoretical safety nets of classical statistics.
  • Guaranteed Performance: Unlike many deep learning approaches, RoS-Guard comes with "proofs in the box." The authors provide a lower bound for the False Alarm Period and an upper bound for the Average Detection Delay, ensuring the system remains reliable under pressure.

Key Results

  • Superior Robustness: In smart grid simulations, RoS-Guard consistently maintained a lower detection delay than state-of-the-art baselines (like RGCUSUM and CyberQCD) when the system model was inaccurate.
  • Massive Speedup: The GPU-accelerated neural unrolling version achieved a 20x speedup over traditional CPU-based solvers as the system dimension scaled up to $2^8$.
  • Theoretical Validation: Experimental results closely tracked the theoretical bounds derived in the paper, confirming that the "relaxed" math didn't compromise the algorithm's fundamental reliability.

Limitations and Future Work

While RoS-Guard is a major step forward, it primarily focuses on linear systems. While many real-world problems can be linearized, extending this robust, unrolled approach to purely non-linear dynamics remains an open challenge. Additionally, while the neural unrolling provides speed, it requires an initial "training" phase to optimize the step-size parameters, which might need to be repeated if the system's broad characteristics change drastically.

Future work could involve exploring even more complex uncertainty sets or integrating the framework with decentralized sensing, where multiple "RoS-Guards" collaborate across a network.

Summary updated

Abstract

Online change detection (OCD) aims to rapidly identify change points in streaming data and is critical in applications such as power system monitoring, wireless network sensing, and financial anomaly detection. Existing OCD methods typically assume precise system knowledge, which is unrealistic due to estimation errors and environmental variations. Moreover, existing OCD methods often struggle with efficiency in large-scale systems. To overcome these challenges, we propose RoS-Guard, a robust and optimal OCD algorithm tailored for linear systems with uncertainty. Through a tight relaxation and reformulation of the OCD optimization problem, RoS-Guard employs neural unrolling to enable efficient parallel computation via GPU acceleration. The algorithm provides theoretical guarantees on performance, including expected false alarm rate and worst-case average detection delay. Extensive experiments validate the effectiveness of RoS-Guard and demonstrate significant computational speedup in large-scale system scenarios.
BibTeX
@misc{zhu2025ros,
  author = {Ze‐Lin Zhu and Yancheng Huang and Kai Yang},
  title = {{RoS-Guard: Robust and Scalable Online Change Detection with Delay-Optimal Guarantees}},
  year = {2025},
  eprint = {2511.12846},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.12846},
  howpublished = {arXiv},
}
Figure