Resilient-by-Design: A Paradigm Shift in Securing Power CPS against False Data Injection and Beyond

1. Introduction

1.1. Motivation: From Reactive Detection to Proactive Resilience

The digitalization of power systems marks a significant leap in grid management capabilities, enabling real-time data acquisition, advanced analytics, distributed control, and predictive maintenance [1,2,3]. Modern power grids have evolved into complex cyber-physical systems (CPS), integrating millions of sensors, intelligent electronic devices (IEDs), renewables, and control systems with sophisticated communication networks [4,5]. This convergence empowers operators to optimize energy production, transmission, and distribution with unprecedented accuracy [6].

However, this transformation also introduces new cyber vulnerabilities [7,8,9]. As critical functions increasingly rely on information and communication technology (ICT), attackers can target not only physical infrastructure but also the digital pathways that control it [10]. Among these emerging threats, False Data Injection Attacks (FDIAs) stand out for their stealth, scalability, and potentially devastating consequences. Unlike conventional cyber-attacks that cause immediate service disruptions, FDIAs silently corrupt system data, deceiving state estimation algorithms and triggering incorrect operational decisions that may lead to cascading failures, blackouts, or equipment damage [11,12,13].

Notably, real-world incidents have already demonstrated the feasibility and impact of cyber-physical attacks on energy infrastructure. The 2015 Ukraine power grid cyberattack caused widespread blackouts affecting over 230,000 people, leveraging a multi-stage attack chain involving malware, remote access, and control manipulation. The Colonial Pipeline ransomware attack in 2021 disrupted fuel supplies across the U.S. East Coast, highlighting the interdependence of energy and cyber infrastructure. These events emphasize that critical infrastructure is no longer protected by physical barriers alone [14,15].

Over the past decade, researchers have made significant progress in developing detection algorithms to identify FDIAs, leveraging techniques such as state estimation residual analysis, machine learning, and data fusion. However, these solutions are largely reactive, focusing on detecting and responding to attacks after they have penetrated the system. This reactive posture has several limitations [16,17,18]:

• Detection delays may allow attacks to inflict damage before mitigation can occur.
• False positives and false negatives can undermine operator trust and lead to inappropriate responses.
• Over-reliance on centralized detection architectures creates bottlenecks and single points of failure.
• Limited scalability and adaptability restrict deployment across large, heterogeneous grid environments.

To break this reactive cycle, the research community and industry must shift toward proactive resilience—designing power CPS that can withstand, absorb, recover from, and adapt to cyber-physical disruptions without catastrophic performance degradation [19]. This concept, known as "Resilient-by-Design", treats resilience not as a patchwork of add-on defenses but as a core design principle embedded throughout the system's lifecycle [20,21,22].

1.2. Scope: What Does “Resilient-by-Design” Mean in Power CPS?

“Resilient-by-Design” represents a systematic, holistic approach to security, emphasizing prevention, detection, response, and recovery as interconnected layers of defense. In the context of power CPS, this paradigm involves [23,24]:

• Secure Sensing and Communication: Ensuring the authenticity, integrity, and availability of measurement and control data through cryptographic protocols, secure communication architectures, and data validation mechanisms.
• Distributed, Real-Time Detection and Response: Empowering local control entities—such as substations, microgrids, and edge devices—with autonomous detection and mitigation capabilities, reducing response latency and improving scalability.
• Data Recovery and Control Reconfiguration: Providing mechanisms for state estimation recovery, topology inference, and control reconfiguration to maintain grid observability and controllability after a cyber compromise.
• Adaptive and Explainable Defense Mechanisms: Leveraging AI-driven algorithms that can learn, adapt, and explain their behavior in evolving threat landscapes, bridging the gap between machine intelligence and operator trust.
• Cross-Domain Coordination: Aligning cyber defense with physical grid operations, regulatory policies, and market mechanisms to ensure resilience extends beyond technical layers to include organizational and societal dimensions.

This review adopts this Resilient-by-Design perspective, integrating insights from cybersecurity, control theory, artificial intelligence, and energy systems engineering. It extends the research frontier beyond FDIA detection, addressing broader questions of systemic resilience, cross-layer defense integration, and operational feasibility.

1.3. Industry and Policy Drivers for Resilient Power CPS

Global policy frameworks increasingly recognize the national security implications of power CPS cyber vulnerabilities [25]. Key initiatives include:

• The North American Electric Reliability Corporation’s Critical Infrastructure Protection (NERC CIP) standards, which mandate baseline cybersecurity practices for U.S. and Canadian utilities.
• The European Union’s Network and Information Systems (NIS2) Directive, expanding cybersecurity obligations for operators of essential services, including energy providers.
• China’s Cybersecurity Law and Critical Information Infrastructure Protection regulations, emphasizing sovereignty and supply chain security in critical sectors.
• The U.S. Department of Energy’s Cybersecurity Capability Maturity Model (C2M2), providing a framework for utilities to assess and improve their cyber resilience.

These policies reflect growing recognition that cyber resilience is a strategic imperative, not merely a technical challenge. Utilities, grid operators, and policymakers must work together to operationalize resilience requirements through:

• Investment in secure infrastructure upgrades.
• Integration of resilience metrics into planning and procurement.
• Cross-sector information sharing and threat intelligence.
• Development of workforce skills in cybersecurity and power engineering.

1.4. Key Contributions and Structure of This Review

This review aims to bridge research, policy, and industry practices, providing a comprehensive roadmap for Resilient-by-Design power CPS. Key contributions include:

Characterization of the evolving threat landscape, highlighting the shift from isolated FDIA to coordinated, multi-vector, and AI-driven attacks.

Framework for security-aware grid architecture design, emphasizing redundancy, diversity, and modularity.

Review of distributed detection and response mechanisms, including federated learning, multi-agent systems, and edge intelligence.

Insights into post-attack data reconstruction and control recovery, essential for maintaining operational continuity.

Exploration of adaptive AI-driven security strategies, including reinforcement learning, generative modeling, and explainable AI.

Identification of future research directions and policy challenges, promoting actionable strategies for cross-sector resilience.

2. Evolution of Power CPS Threat Landscape

2.1. From Classic FDIA to Coordinated Multi-Stage Attacks

The first formal definition of FDIA by Liu et al. in 2009 was a landmark contribution to power system cybersecurity [27]. It demonstrated that attackers could mathematically craft stealthy false data that aligns with the physical constraints of power flow models, making it undetectable by conventional bad data detection based on residual checks. This inspired a wave of research focusing on state estimation security, laying the foundation for cyber-physical security as a research discipline in power systems [28,29,30].

Limitations of Early FDIA Models

However, early models often assumed full system knowledge and unrestricted access to measurements—assumptions that may not hold in real-world attacks [31,32]. In practice, adversaries may face partial information, limited access, and stochastic network conditions. Despite these practical barriers, real-world incidents have shown that attackers do not need perfect information to cause significant disruption [33].

Transition to Multi-Stage and Coordinated Attacks

Over the past decade, attack strategies have evolved from isolated data injection toward multi-stage, coordinated campaigns that combine [34]:

• Initial reconnaissance, such as scanning network ports or phishing operator credentials.
• Lateral movement, compromising multiple devices or systems to escalate privileges.
• Persistence, embedding malware to maintain long-term system access.
• Coordinated manipulation, simultaneously targeting measurement data, control signals, and operator interfaces.

These attacks often leverage supply chain vulnerabilities, insider threats, or third-party software exploits. The 2015 and 2016 Ukraine grid attacks serve as prime examples, where attackers combined malware (BlackEnergy and Industroyer), remote access tools, and manual control room manipulation to orchestrate widespread outages.

Adversarial Goals Beyond Outages

Modern cyber-physical adversaries may pursue objectives beyond service disruption, including [35,36,37]:

• Economic manipulation, such as exploiting electricity markets by causing price spikes or imbalance penalties.
• Physical equipment damage, as seen in Stuxnet’s targeted sabotage of centrifuges.
• Erosion of operator trust, causing hesitation or incorrect responses during critical events.
• Long-term espionage, gathering data on system operations, vulnerabilities, or market behaviors.

Understanding these multi-dimensional attack objectives is crucial for developing defense strategies that go beyond availability protection, addressing integrity, confidentiality, and trust.

2.2. Emerging Attack Vectors: AI-Driven Attacks, Federated Poisoning, Time-Sync Attacks

AI-Driven Adversarial Strategies

As defenders increasingly adopt machine learning (ML) and artificial intelligence (AI) for anomaly detection, attackers are adapting by using AI themselves to optimize their attacks. Adversarial machine learning techniques enable attackers to [38,39,40]:

• Train surrogate models of the grid based on observed data.
• Craft adversarial examples that exploit the detection model’s blind spots.
• Deploy adaptive attacks that evolve in response to defense updates.

For example, an attacker might simulate different FDIA patterns using reinforcement learning to identify the most stealthy and effective manipulation strategies.

Federated Learning Poisoning Risks

Federated learning allows distributed agents to collaboratively train models without sharing raw data, improving privacy [41,42,43]. However, this opens new attack surfaces, such as:

• Model poisoning, where compromised agents submit malicious updates to degrade detection performance.
• Sybil attacks, where a single adversary controls multiple seemingly independent agents.
• Backdoor insertion, embedding hidden triggers in the global model.

Mitigating these risks requires robust aggregation techniques (e.g., Krum, Bulyan) and participant verification mechanisms.

Time Synchronization and GPS Spoofing

Time synchronization is critical for PMU-based state estimation [44]. GPS spoofing attacks can introduce:

• Timestamp misalignment, leading to inaccurate state estimation.
• Inter-PMU desynchronization, degrading the performance of wide-area monitoring systems.
• Control instability, if protection or control schemes rely on precise timing.

Given the low-cost and accessible nature of GPS spoofers, this attack vector poses a significant operational risk, especially in wide-area or islanded microgrid applications [45,46,47].

Advanced Control Hijacking and Data-Driven Control Manipulation

Adversaries may go beyond data corruption to manipulate control logic itself, for instance by [48]:

• Exploiting weakly secured remote terminal units (RTUs) to inject malicious setpoints.
• Manipulating distributed energy resource (DER) controllers to destabilize frequency or voltage.
• Targeting energy management systems (EMS) to disrupt optimal dispatch or market operations.

These attacks are difficult to detect if they mimic legitimate operator behavior or exploit trust relationships between devices and control centers.

2.3. Cross-Domain Attack Modeling: ICT-OT Coupling Dynamics

The Need for Cross-Domain Security Models

Power CPS are not isolated from other infrastructures. They rely on telecommunications, IT services, and cloud platforms, creating cross-domain dependencies that adversaries can exploit [49,50].

For instance:

• Telecom network failures can block operator communications during cyber-physical events.
• Cloud service compromises may expose configuration or credential data.
• Third-party vendors providing SCADA or EMS software may introduce supply chain risks.

Example: Coordinated Attack on Energy and Telecom

Consider a hypothetical scenario where an attacker:

(1)

Launches a DDoS attack on the telecom provider’s network, delaying operator situational awareness.

(2)

Simultaneously injects false data into substation measurement streams.

(3)

Exploits compromised VPN credentials to access control systems.

Such cross-layer, multi-domain attacks require defense mechanisms that integrate cyber and physical situational awareness, bridging the gap between ICT and OT domains [51,52,53].

Modeling Multi-Layered Propagation Effects

Advanced simulation tools, such as co-simulation platforms that link power system models with communication network models, are essential to [54,55]:

• Analyze propagation effects of cross-domain attacks.
• Evaluate cascading risks across ICT and OT layers.
• Design coordinated defense strategies that account for multi-domain interdependencies.

2.4. Summary of Threat Evolution

Threat Evolutions are summarized in Table 1.

3. Design Foundations for Resilient Power CPS

Building resilience into power CPS requires rethinking system architecture, control strategies, and operational workflows. Instead of retrofitting security onto legacy systems, a Resilient-by-Design approach integrates defense mechanisms from the ground up, balancing operational efficiency, security assurance, and scalability [56,57,58].

3.1. Security-Aware Grid Architecture: Redundancy, Diversity, and Modularity

Redundancy: Engineering Out Single Points of Failure

Redundancy is a well-known principle in power system reliability engineering, traditionally applied to physical infrastructure such as transmission lines, transformers, and protection relays. However, cyber redundancy—including sensing, communication, and control redundancy—is equally critical in mitigating cyber-physical threats [59,60].

• Redundant Sensing: Deploying multi-source measurements (e.g., PMUs, RTUs, and smart meters) provides cross-verification capabilities, reducing reliance on any single data stream.
• Redundant Communication: Establishing diverse communication paths, such as wired fiber networks and wireless LTE/5G links, ensures that data can flow even if one channel is disrupted.
• Redundant Control Logic: Implementing parallel control algorithms, possibly on independent hardware or software platforms, provides fallback options if one control path is compromised.

The goal is to localize the impact of cyber-physical disruptions and enable graceful degradation rather than total system collapse [61,62,63].

Diversity: Breaking Homogeneity to Increase Attack Complexity

Homogeneous systems—those built on identical hardware, software, and protocols—present an attractive target for attackers, who can generalize their exploits across the entire infrastructure. Diversity, in contrast, increases the cost and complexity of attacks by forcing adversaries to navigate heterogeneous defense surfaces [64,65].

• Vendor Diversity: Using equipment from multiple manufacturers reduces the risk of a single vendor vulnerability affecting the entire system.
• Software Diversity: Running different firmware versions or diverse operating systems mitigates the impact of zero-day exploits.
• Protocol Diversity: Supporting multiple communication protocols (e.g., IEC 61850, DNP3, Modbus) with gateway isolation can prevent protocol-specific attacks from spreading system-wide.

However, managing diversity introduces operational challenges, such as interoperability management and training requirements. These must be balanced against the security benefits through risk-informed engineering trade-offs [66,67,68,69].

Modularity: Isolating Risks through System Segmentation

Modularity involves partitioning the grid into semi-independent, self-sufficient segments with clear trust boundaries [70,71,72]. Examples include:

• Microgrids capable of islanded operation during cyber or physical disturbances.
• Virtual Power Plants (VPPs) aggregating DERs with autonomous control capabilities.
• Regional Control Zones with localized monitoring and response mechanisms.

Modular architectures enable localized resilience, reducing the risk of cascading failures across the entire grid. They also support incremental deployment of advanced security features without requiring a complete system overhaul [73].

Supply Chain and Physical Security Integration

True Resilient-by-Design architectures must also consider [74,75]:

• Supply chain security, ensuring that hardware and software components are free from embedded vulnerabilities.
• Physical security integration, protecting critical cyber-physical assets from tampering, sabotage, or insider threats.

This requires end-to-end security assurance, covering procurement, installation, operation, and decommissioning phases [76,77].

3.2. Trustworthy Sensing and Communication: Secure-by-Protocol vs. Secure-by-Learning

Secure-by-Protocol: Cryptographic Protection of Data Integrity

Protocol-level security remains foundational for defending against data manipulation and eavesdropping. Key mechanisms include [78,79,80]:

• Message Authentication Codes (MACs): Verifying that messages have not been altered in transit.
• Digital Signatures: Providing non-repudiation and source authentication.
• End-to-End Encryption: Protecting data confidentiality from source to destination.

Standards such as IEC 62351 provide guidelines for securing SCADA and substation communications. However, practical challenges include [81]:

• Key management complexity in large, distributed systems.
• Computational overhead on resource-constrained edge devices.
• Legacy system limitations, where older devices lack cryptographic capabilities.

Secure-by-Learning: Data-Driven Trust Validation

Complementing protocol-level security, data-driven validation mechanisms can detect:

• Statistical anomalies in measurement patterns.
• Physical model inconsistencies, such as violations of power flow equations.
• Temporal or spatial deviations from normal system behavior.

Machine learning models, including physics-informed neural networks, can learn normal operational baselines and flag deviations that may indicate cyber-physical manipulation. These methods are particularly valuable when [82,83,84]:

• Cryptographic protection is infeasible or compromised.
• Anomalies bypass signature-based detection.

However, model generalizability, explainability, and resilience to adversarial manipulation remain active research challenges [85].

Hybrid Secure-By-Design Integration

The most resilient architectures combine secure-by-protocol and secure-by-learning approaches, providing defense-in-depth. For example:

• Cryptographic mechanisms ensure data authenticity at the transport layer.
• Anomaly detection models validate data plausibility at the application layer.

This layered defense ensures that even if one protection mechanism is bypassed, complementary safeguards remain in place [86,87,88].

3.3. Control-Theoretic Resilience Metrics and Stability under Adversarial Perturbation

Quantifying Cyber-Physical Robustness

Control-theoretic resilience metrics provide a quantitative foundation for designing systems that can tolerate bounded cyber-physical disturbances [89,90]. Key metrics include:

• Input-output stability margins: Measuring the system's ability to absorb input disturbances without destabilizing output responses.
• Region of attraction under uncertainty: Defining the safe operating region despite model uncertainties or data corruption.
• Resilient controllability and observability: Ensuring that critical states remain observable and controllable even if some data channels are compromised.

Stability-Aware Control Algorithm Design

Resilient control algorithms can be designed using [91,92,93]:

• Robust control techniques (e.g., H-infinity control) that minimize worst-case impacts of bounded disturbances.
• Adaptive control methods that adjust controller parameters in response to detected anomalies.
• Sliding mode control that enforces stability through discontinuous control actions, effective against model uncertainties.

Graceful Degradation and Emergency Protocols

In severe attack scenarios, maintaining partial functionality may be preferable to total shutdown. Graceful degradation strategies include [94,95]:

• Prioritizing critical loads (e.g., hospitals, data centers).
• Selective islanding of microgrids.
• Fallback to manual control when automation is compromised.

Pre-defined emergency control protocols, validated through simulation and field exercises, ensure that operators know when and how to transition to degraded modes safely.

4. Real-Time Distributed Detection and Response Mechanisms

4.1. Federated and Privacy-Preserving FDIA Detection

The growing scale and geographical distribution of power CPS make centralized detection architectures increasingly impractical. Centralized methods face scalability challenges, high communication overhead, and data privacy concerns. To address these limitations, federated learning has emerged as a distributed and privacy-preserving paradigm for collaborative FDIA detection [96,97,98].

In federated learning, local agents (e.g., substations, microgrids, or edge devices) train detection models on locally available data without sharing raw data with a central server. Only model updates or gradients are exchanged and aggregated to form a global model. This approach offers several advantages:

• Data privacy preservation, as raw measurements never leave local devices.
• Reduced communication overhead, since model updates are typically smaller than raw datasets.
• Scalability, enabling large-scale deployments across geographically distributed infrastructures.

However, federated learning also introduces new challenges, including model poisoning attacks, where malicious participants submit deceptive updates to degrade global model performance. To mitigate this risk, robust aggregation algorithms and participant reputation mechanisms must be integrated into the federated learning process [99,100].

4.2. Multi-Agent-Based Cooperative Defense Strategies

Power CPS inherently consist of interconnected yet semi-autonomous subsystems, such as regional control centers, microgrids, and substations [101,102,103]. These subsystems can be modeled as intelligent agents capable of:

• Local detection, based on their own measurements and historical data.
• Cooperative information sharing, exchanging detection results, alerts, or suspicious patterns with neighboring agents.
• Distributed decision-making, coordinating response actions to contain or mitigate detected threats.

Multi-agent systems enable cooperative defense without over-relying on a single point of control. For example, if one agent detects anomalous behavior, it can propagate warnings to its neighbors, enabling collective situational awareness [104,105,106]. This peer-to-peer defense mechanism enhances resilience against stealthy attacks that target isolated subsystems.

Effective multi-agent cooperation requires [107,108]:

• Consensus protocols to resolve conflicting assessments among agents.
• Trust management frameworks to evaluate the credibility of information sources.
• Secure communication channels to prevent attackers from hijacking agent interactions.

By leveraging the distributed nature of power CPS, multi-agent systems can provide scalable, robust, and adaptive defense capabilities.

4.3. Edge Intelligence and On-Device Learning for Local Anomaly Detection

As computational resources become increasingly available at the network edge (e.g., in substations or IoT devices), edge intelligence offers a promising solution for real-time, localized FDIA detection [109,110].

Edge intelligence involves deploying lightweight machine learning models or streaming analytics engines directly on edge devices, enabling them to [111,112,113]:

• Continuously monitor local data streams (e.g., voltage, current, frequency).
• Detect anomalies in real time, without relying on centralized processing.
• Trigger immediate local responses, such as isolating compromised components or raising alarms.

Key enablers of edge intelligence include [114]:

• Model compression and resource-aware learning algorithms, which reduce the computational footprint of machine learning models.
• Online and incremental learning, allowing models to adapt to evolving grid conditions without full retraining.
• Explainable AI techniques, providing interpretable detection results to support operator trust and human-in-the-loop decision-making.

By shifting detection and response capabilities closer to the data sources, edge intelligence reduces detection latency, improves scalability, and enhances system autonomy, making it a critical component of resilient power CPS defense architectures.

5. Data Reconstruction and Recovery after Compromise

While detection and isolation are crucial first steps in cyber defense, what happens after an attack is equally critical. Simply discarding compromised data or disconnecting affected assets can leave the grid in a degraded or unobservable state, potentially leading to secondary failures or operational blind spots. Therefore, resilient power CPS must include data reconstruction and system recovery capabilities to safely restore operations [115,116,117].

5.1. Resilient State Estimation and Control Recovery

While much of the existing literature focuses on detecting FDIAs, less attention has been given to what happens after detection [118,119,120]. In practice, simply flagging or discarding suspected data may not be sufficient, especially if large portions of measurement data are compromised. This could lead to loss of observability and unreliable control actions [121].

Resilient state estimation addresses this challenge by enabling the power system to:

• Reconstruct missing or corrupted states using redundant measurements, historical data, or predictive models.
• Estimate confidence levels for reconstructed states, supporting risk-informed operational decisions.
• Reconfigure control strategies based on partially trusted data, prioritizing stability and safety.

Advanced techniques, such as moving horizon estimation, robust Kalman filtering, and physics-informed neural networks, have shown potential for enhancing state estimation resilience under partial data loss. These methods combine physical system models with data-driven insights, enabling adaptive estimation even in degraded sensing environments [123,124,125].

5.2. Topology and Data Restoration after Multi-Point Compromise

Multi-point FDIAs often target both measurement data and network topology information, making recovery particularly challenging. Effective topology restoration requires [126,127,128]:

• Cross-validation of topology information using multiple independent data sources (e.g., PMU data vs. SCADA data).
• Topology inference algorithms, capable of reconstructing the most likely network structure based on surviving measurements and physical constraints.
• Anomaly-tolerant network reconfiguration, which can isolate suspicious areas while maintaining system-wide observability.

For example, if an attacker manipulates breaker status data to hide the true network topology, data-driven topology estimation methods can help reconstruct the likely physical configuration, reducing the risk of operating on false assumptions [129].

Furthermore, redundant data pathways, such as alternative communication channels or secondary measurement systems, can provide fallback information to support post-attack recovery. Leveraging such redundancy is critical to restoring situational awareness after a coordinated attack [130].

5.3. Redundant Data Pathways and Confidence-Aware Data Fusion

Resilient recovery also relies on fusing data from multiple, potentially imperfect sources. Confidence-aware data fusion techniques aim to [131,132,133]:

• Quantify uncertainty in each data source, based on historical reliability, detection results, or model consistency checks.
• Combine multiple data streams in a weighted manner, giving higher priority to more trusted sources.
• Provide operators with confidence scores, supporting informed decision-making under uncertainty.

For instance, if PMU data appears suspicious due to detected time-sync attacks, the system can downgrade its confidence in that data and rely more on SCADA or historical trend data. This adaptive trust management helps maintain operational awareness even when some data sources are compromised [134].

Additionally, blockchain-based data provenance and secure logging mechanisms can help track the integrity and origin of measurement data, providing further assurance during recovery processes.

In summary, resilient power CPS require not only detection but also intelligent recovery mechanisms that leverage redundancy, cross-validation, and confidence-aware fusion to restore safe and reliable operations after cyber-physical compromises [135,136].

6. Autonomy-Driven Adaptive Security: From Rules to Reasoning

Conventional cyber defense for power CPS often relies on static security policies, fixed detection thresholds, and manual operator intervention. While these methods provide basic protection, they fail to scale in the face of dynamic, evolving, and adaptive adversaries [137]. Moreover, overly rigid defenses risk generating excessive false positives, overwhelming operators and degrading trust in security systems. To address these limitations, autonomy-driven adaptive security leverages artificial intelligence (AI) and machine learning (ML) to enable self-learning, self-adaptive, and self-explainable defense mechanisms [138,139,140]. This section explores cutting-edge techniques that transform static defenses into reasoning-capable, autonomous security agents.

6.1. Reinforcement Learning for Adaptive Cyber Response

Traditional cyber defense mechanisms often rely on static rule sets or pre-defined thresholds, which struggle to cope with the evolving and dynamic nature of cyber-physical attacks. Reinforcement learning (RL) provides a promising paradigm for adaptive cyber response, enabling power CPS to learn optimal defense strategies through interaction with the environment [141,142,143].

In an RL framework:

• The power CPS is modeled as an environment with states (e.g., grid measurements, detection alerts), actions (e.g., isolating a node, switching control modes), and rewards (e.g., maintaining stability, minimizing false positives).
• An agent learns a policy that maps observed states to actions that maximize long-term resilience [144].

Key advantages of RL-based defense include:

• Adaptability to unseen attack patterns or evolving grid conditions.
• Autonomous policy improvement through continuous learning.
• Balance between false alarms and missed detections, optimizing operational impact.

Recent research has demonstrated the feasibility of deep reinforcement learning (DRL) for cyber-physical defense, leveraging neural networks to handle high-dimensional state spaces [145,146,147]. However, ensuring safe exploration, policy interpretability, and real-time convergence remain open challenges for practical deployment [148].

6.2. Generative Models for Anticipatory Defense

While detection and response are reactive by nature, anticipatory defense aims to predict and preempt attacks before they fully materialize [149,150]. Generative models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), offer tools for [151,152,153,154]:

• Simulating potential attack scenarios to stress-test defense mechanisms.
• Generating synthetic attack data to improve detection model robustness.
• Forecasting likely system states under adversarial conditions, enabling preemptive mitigation.

For example, GAN-based frameworks can generate plausible false data injection patterns, allowing the defender to train detection models on a wider variety of attack vectors. Similarly, VAEs can model normal system behavior, flagging deviations that might indicate stealthy attacks [155,156,157,158,159,160].

By integrating generative models into the cyber defense lifecycle, power CPS can proactively explore the adversarial space, improving preparedness against sophisticated and adaptive attackers.

6.3. Explainable AI for Operator Trust and Human-in-the-Loop Security

As AI and machine learning increasingly drive cyber defense, explainability becomes critical for operator trust and effective human-machine collaboration [161,162,163]. Explainable AI (XAI) aims to make AI-driven decisions transparent, interpretable, and actionable for human operators [164,165,166,167,168].

Key aspects of XAI in power CPS defense include [169,170]:

• Visualizing detection rationale, such as highlighting which sensors or features contributed most to a detection decision.
• Providing confidence scores and uncertainty estimates, supporting risk-informed operator actions [171].
• Enabling what-if analyses, allowing operators to explore the implications of different response strategies [172].

Human-in-the-loop security frameworks combine AI automation with human judgment, ensuring that high-consequence decisions (e.g., grid reconfiguration, load shedding) are informed by both machine insights and operator expertise [173,174,175].

Ultimately, explainable and human-centered AI enhances trust, accountability, and operational reliability, ensuring that advanced defense mechanisms are not black boxes but transparent allies in securing power CPS [176,177,178].

7. Future Directions and Open Challenges

7.1. Resilience Co-Design with Grid Planning and Market Mechanisms

Most existing cyber defense research treats security as a runtime feature, decoupled from long-term grid planning and market design [179,180,181]. However, true resilience requires integrating security objectives into infrastructure planning, investment strategies, and market operations [182,183].

Future work should explore:

• Co-optimization of security and economic objectives, ensuring that resilience measures do not undermine market efficiency.
• Incentive mechanisms for distributed energy resources and microgrids to participate in grid-wide cyber defense, leveraging their local intelligence and control capabilities.
• Security-aware planning models that account for the costs and benefits of redundant infrastructure, diverse vendor ecosystems, and modular grid segmentation.

By embedding resilience into the economic and regulatory fabric of power systems, operators can move beyond reactive defense toward proactive, economically justified security investments [184,185,186].

7.2. Adversarial Testing and Red-Teaming Frameworks

Current defense solutions are often validated under limited, idealized scenarios. There is a pressing need for systematic adversarial testing to evaluate resilience under realistic, worst-case attack conditions [187,188,189,190].

Future research should focus on:

• Red-teaming frameworks, where expert adversaries simulate advanced cyber-physical attack strategies against defense mechanisms [191].
• Adversarial AI techniques to stress-test detection models, revealing blind spots and robustness limits [192,193,194].
• Digital twin environments, providing safe, high-fidelity simulation platforms for end-to-end security validation without risking real-world operations.

Such testing frameworks can build confidence in proposed solutions and accelerate their transition from academic prototypes to operational deployment [195].

7.3. Benchmarking and Datasets for Realistic Evaluation

The lack of standardized datasets and benchmarking frameworks remains a major bottleneck in advancing power CPS security research [196,197,198]. Many studies rely on synthetic data or simplified test systems, limiting the generalizability of their findings [199,200,201].

To address this, the community needs [202,203,204,205]:

• Open, realistic datasets, including multi-domain (cyber-physical) measurements, labeled attack scenarios, and diverse operating conditions.
• Standardized evaluation metrics, enabling fair comparison of detection accuracy, response latency, scalability, and resilience impact.
• Shared benchmarking platforms, where researchers can test their methods under consistent and challenging conditions.

Collaboration among academia, industry, and government is essential to curate and maintain these resources, ensuring reproducibility and comparability across studies [206,207,208,209,210].

7.4. Policy, Standardization, and Cross-Sector Collaboration

Technical advancements alone are insufficient to secure power CPS [211,212,213]. Policy, governance, and cross-sector collaboration are equally critical [214].

Key priorities include [215,216,217]:

• Establishing security standards and compliance frameworks for power CPS, covering sensing, communication, control, and data management layers.
• Facilitating information sharing among utilities, manufacturers, cybersecurity experts, and regulators to accelerate threat intelligence dissemination and best practice adoption.
• Promoting cross-sector resilience planning, recognizing that power systems are interdependent with telecommunications, transportation, and other critical infrastructures.

Building a trusted ecosystem of stakeholders, supported by clear policies and standards, is vital to achieving systemic, cross-domain resilience in the face of evolving cyber threats [218].

8. Conclusions

The evolution of power systems into complex cyber-physical systems has unlocked unprecedented operational capabilities but has also exposed the grid to a rapidly expanding range of cyber-physical threats. Among these, false data injection attacks have emerged as particularly stealthy and damaging, capable of undermining grid stability and operator trust.

While substantial progress has been made in FDIA detection, this review highlights the urgent need to move beyond reactive defense toward a Resilient-by-Design paradigm. Such a paradigm reimagines power CPS security as a system-level, co-designed capability, embedded across sensing, communication, control, and recovery layers.

Key takeaways from this review include:

• The evolving threat landscape now extends beyond classic FDIAs to include coordinated multi-stage, AI-driven, and cross-domain attacks.
• Security-aware grid architectures, emphasizing redundancy, diversity, and modularity, form the foundation for systemic resilience.
• Distributed, real-time defense mechanisms, leveraging federated learning, multi-agent systems, and edge intelligence, offer scalable and adaptive protection.
• Data reconstruction and recovery are essential for maintaining situational awareness and operational continuity after an attack.
• Autonomy-driven adaptive security, powered by reinforcement learning, generative modeling, and explainable AI, enhances the system’s ability to learn, adapt, and build operator trust.

Looking ahead, achieving practical, large-scale deployment of resilient power CPS requires:

• Integrating resilience objectives into grid planning and market mechanisms.
• Systematic adversarial testing and red-teaming to validate defense strategies under realistic conditions.
• Standardized datasets and benchmarking frameworks to enable reproducibility and comparability of research findings.
• Cross-sector collaboration and policy alignment to foster a trusted, resilient energy ecosystem.

By embracing these directions, the power systems community can transition from a fragmented, detection-centric security posture to a holistic, proactive, and resilient defense framework, capable of safeguarding critical infrastructure in an increasingly adversarial cyber-physical landscape.