Over the past two centuries, electric power grids have become a cornerstone of modern society. Virtually every critical service depends on reliable electricity, from healthcare and communications to transportation and financial systems. 

Critical infrastructure systems are growing more complex, expanding the attack surface and increasing security challenges. Recurring outages and concerns about grid resilience highlight their vulnerability, particularly to deliberate attacks by terrorist groups or nation-state adversaries. Understanding emerging threats and developing effective countermeasures is essential.

At the 2026 IEEE Conference on Control Technology and Applications (CCTA), researchers examine feedback instability attacks on power grids, where an attacker introduces unstable dynamics that can trigger cascading outages. The researchers present a novel, data-driven enhancement to existing load-shedding-based remedial action schemes. 

Proposed Mitigation Framework

The proposed approach enables operators to detect and mitigate instability attacks before significant damage occurs and provides attack-detection methods that let operators train models on their own systems and deploy software-based response logic, offering a practical, cost-effective defense against this emerging class of cyberattacks.

Consider an instability attack in progress:

  • Modified Prony Analysis (MPA), traditionally used for grid stability monitoring, detects the abnormal dynamics and triggers a response. 
  • The system then passes measurements to a machine-learning model, deployed at the edge or in the cloud, that identifies loads whose shedding is likely to stabilize the system and those that may worsen the attack. 
  • An operator, or automated controller, selects an action based on these recommendations. 
  • If the operator sheds an appropriate load, the instability is suppressed, and normal grid operation can be restored.

For each attack scenario and candidate load, the researchers simulated the shedding action and recorded system states up to a well-defined pre-shedding point. The resulting dataset consists of the candidate load, the pre-shedding system state, and a label indicating whether shedding that load leads to a stable or unstable outcome. This dataset forms the basis for training the machine-learning model.

An example of naively shedding according to under-voltage load-shedding schemes under an instability attack.

 

Strengths: Operational Flexibility and Low-Cost Deployment

The framework uses supervised machine learning to bypass many challenges of analytical approaches, including computational complexity, system identification, and real-time decision-making constraints. By learning directly from system-state data, the model can rapidly identify effective mitigation actions during an attack.

Rather than recommending a single load for shedding, the model classifies all candidate loads as either stable-shed or unstable-shed options. This provides operators with operational flexibility, allowing them to incorporate additional considerations such as customer impact, load criticality, system recovery objectives, and other forms of stability beyond the small-signal stability targeted by instability attacks.

This approach is also cost-effective. Unlike mitigation strategies that require physical modifications to grid topology or infrastructure, the proposed framework is software-based and builds upon existing monitoring, protection, and contingency-management systems. As a result, it can be deployed with comparatively low implementation costs.

Weaknesses: Data Availability, Latency, and Integrity Challenges

Although cloud-connected industrial control systems are becoming more common, the current framework assumes access to measurements across the entire power system. This can be challenging in systems that contain legacy equipment. Furthermore, conventional SCADA systems often lack the temporal resolution needed to respond effectively to instability attacks, whose dynamics may evolve much faster than standard monitoring intervals.

Because power grids are geographically distributed, delays in transmitting measurements to centralized control centers may hinder rapid response. In practice, mitigation actions may need to be executed locally before being reported to a central authority. Extending this framework to decentralized or multi-agent control architectures represents a promising direction for future work.

Finally, the framework depends on the integrity of sensor data. If an adversary can spoof or manipulate measurements before they reach the detection or decision-making systems, the resulting system-state estimates may be inaccurate, reducing the effectiveness of the proposed mitigation strategy. Ensuring trustworthy data collection and communication is therefore critical to successful deployment.

Toward Practical Mitigation of Instability Attacks

Although instability attacks are not yet widespread, the increasing targeting of critical infrastructure in modern cyber warfare suggests that such threats may become more common in the future. Developing effective mitigation strategies before they emerge at scale is therefore essential.

The researchers posit that conventional load-shedding strategies can be ineffective or even harmful during instability attacks, motivating the need for more adaptive approaches. This paper establishes a foundation for scalable, industry-grade tools that can defend critical infrastructure against instability attacks, providing a pathway to practical deployment in real-world power systems.

Interested in learning more about Grid Defense and Machine Learning? We have thousands of articles related to these industries and more! Also, learn more about our eBooks and eLearning collections on cutting-edge technologies.

Interested in acquiring full-text access to this collection for your entire organization? Request a free demo and trial subscription for your organization.