⚡ electrical engineering

Byzantine-Resilient Federated Multi-Agent Optimization Framework for Cyber-Secure Interconnected Microgrids

This paper proposes BR-FedMAPPO, a Byzantine-resilient federated multi-agent reinforcement learning framework that secures interconnected microgrids against stealthy false data injection attacks by combining privacy-preserving model partitioning, adaptive isolation strategies, and a two-stage aggregation rule to mitigate cyber threats while maintaining cost-effective dispatch.

Original authors: Ali Peivand, Seyyed Mostafa Nosratabadi

Published 2026-08-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Peivand, Seyyed Mostafa Nosratabadi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid not as a giant, silent machine, but as a bustling city of tiny, independent neighborhoods called microgrids. Each neighborhood has its own power generators, batteries, and lights, but they are all connected by roads (power lines) so they can share energy when one is running low. In the past, these neighborhoods were simple and closed off. Today, they are super-smart and connected to the internet, allowing them to talk to each other and balance power automatically. This is great for efficiency, but it opens a door for digital thieves. These thieves don't break in with crowbars; they use "Stealthy False Data Injection" attacks. Think of it like a hacker whispering a lie to a traffic camera, making it think a road is clear when it's actually blocked. Because the lie is so perfect, the city's security system (the Bad Data Detector) doesn't even notice, and the traffic (electricity) gets sent into a dangerous crash.

To stop these invisible hackers, scientists have been trying two main things. First, they try to build better locks (detection), but hackers keep finding new ways to pick them. Second, they try to use "Moving Target Defense," which is like a game of musical chairs where the rules of the game change constantly so the hackers can't memorize the board. However, most of these solutions require a single, all-powerful boss to control everything, which violates the privacy of the neighborhoods. They don't want to share their secret blueprints with a central authority. This paper steps in with a clever new idea: a team of neighbors learning together without ever showing their secret plans to each other, while also being smart enough to ignore a neighbor who is trying to trick the group.

The authors of this paper, Ali Peivand and Seyyed Mostafa Nosratabadi, propose a system called BR-FedMAPPO. It's a mouthful, but you can think of it as a "Byzantine-Resilient Federated Multi-Agent" team. Let's break that down. "Federated" means the neighborhoods learn together but keep their private data (like where their batteries are) to themselves. "Multi-Agent" means every neighborhood has its own little robot brain (an AI agent) making decisions. "Byzantine-Resilient" is the coolest part: it means the system can handle a traitor. In the world of these neighborhoods, a "Byzantine" failure is when one neighborhood gets hacked or goes crazy and tries to send bad instructions to the whole group to crash the party. The system is designed to spot that traitor and ignore them, keeping the rest of the team safe.

Here is how their "super-team" works. Instead of just one big brain, every microgrid has its own local brain. These brains share a common "understanding" of the world (a shared encoder) but keep their own specific "action plans" (private heads) secret. This way, they learn from each other's experiences without revealing their unique layouts. The robot brains have a triple-threat strategy to confuse hackers:

Wiggling the Wires: They slightly change the resistance of the power lines (using devices called D-FACTS), making the electrical map shift constantly.
Moving the Batteries: They change how fast the neighborhood batteries charge or discharge, altering the energy signature.
Changing the Flow: They adjust how much power is sent to neighbors, disrupting any pattern a hacker might be trying to predict.

If a hacker tries to inject false data based on the old map, the map has already changed, and the lie is instantly exposed. But what if a hacker takes over one neighborhood's brain and tries to poison the group's learning? That's where the "Reward-Weighted" rule comes in. The central server doesn't just average everyone's advice. Instead, it looks at how well each neighborhood is doing. If a neighborhood starts raising false alarms (saying "Attack!" when there is none) or acting strangely, the system automatically lowers its voting power. It's like a group project where if one student keeps submitting wrong answers, the teacher stops counting their grade toward the final score.

The researchers tested this idea in a computer simulation using two famous electrical grid models: a smaller one with 30 connection points (buses) and a larger one with 118. They simulated a scenario where hackers launched coordinated, sneaky attacks. The results were promising. The new system successfully stopped the hackers from hiding their tracks, catching about 95% of the coordinated bursts of bad data. It also showed that the system could handle a "traitor" neighborhood trying to mess up the learning process, reducing the error caused by such attacks by a massive 98.62% compared to standard methods.

One of the most interesting features is the "Adaptive Islanding." If the system detects that a neighborhood is truly compromised and the threat is too big to handle, it can automatically "island" itself. This is like a neighborhood cutting its power lines to the rest of the city to prevent the fire from spreading. The paper shows that this happens intelligently; the system doesn't panic and cut the power for every little glitch. It waits until it's sure, and when it does cut the connection, it keeps the neighborhood stable on its own. The simulations showed that this method prevented the chaos from spreading to neighboring towns, keeping their voltage stable and their lights on.

The paper also highlights a trade-off. While the system is very good at catching the big, coordinated lies, it is less sensitive to tiny, random noise. The authors explain this is a feature, not a bug. If the system tried to catch every single tiny glitch, it would start panicking and cutting power unnecessarily, which would be expensive and annoying. Instead, it focuses on the big threats. The simulations suggest that this approach keeps the cost of running the grid low while keeping it safe. For instance, in the smaller test system, the new method actually lowered operational costs for some neighborhoods by nearly 40% compared to older methods, proving that being secure doesn't have to mean being expensive.

In short, this paper presents a simulation of a smart, cooperative defense system for power grids. It shows that by letting neighborhoods learn together without sharing secrets, and by having a built-in mechanism to ignore the "bad apples" in the group, we can create a grid that is both private and incredibly hard to hack. The authors suggest that this "triple-surface" defense—changing wires, batteries, and flows all at once—makes it nearly impossible for hackers to stay hidden. While these results are currently just computer simulations, they offer a very strong blueprint for how we might protect our future, interconnected power systems from the digital threats of tomorrow.

Technical Summary: BR-FedMAPPO for Cyber-Secure Interconnected Microgrids

1. Problem Statement

The rapid digitalization of distribution networks has transformed interconnected Microgrid (MG) clusters into complex cyber-physical systems vulnerable to Stealthy False Data Injection Attacks (S-FDIAs). Unlike random noise, these attacks are crafted to bypass conventional Bad Data Detectors (BDDs) by manipulating state estimation without triggering residual tests. The threat is exacerbated in interconnected MGs, where a compromised unit can propagate corrupted measurements through tie-lines and shared control algorithms.

Existing defenses face three primary limitations:

Centralized Control: Traditional Moving Target Defense (MTD) strategies often rely on centralized coordination, violating the privacy constraints of multi-owner MG clusters.
Static Defenses: Classical detection schemes are inadequate against adversaries who train surrogate models on historical data to bypass deep-learning detectors.
Federated Vulnerabilities: While Federated Learning (FL) offers privacy-preserving collaboration, standard aggregation rules (e.g., Federated Averaging) are susceptible to Byzantine faults, where a single compromised participant can derail the global policy. Furthermore, current Federated Multi-Agent Reinforcement Learning (MARL) approaches often ignore physical coupling between MGs and lack mechanisms for adaptive islanding or structural privacy regarding device placement.

2. Methodology: BR-FedMAPPO

The paper proposes BR-FedMAPPO, a Byzantine-Resilient Federated Multi-Agent Proximal Policy Optimization framework. The system models the cooperative defense problem as a partially observable Markov game where each MG hosts a local Actor-Critic agent.

Core Architectural Components:

Triple-Surface Moving Target Defense (MTD): The action space is partitioned into three physical control surfaces to dynamically alter the attack surface:
1. Reactance Control: Perturbing D-FACTS line reactances to shift the measurement Jacobian, rendering pre-calculated attack vectors invalid.
2. Storage Dispatch: Redirection of Battery Energy Storage (BES) injections to alter power-injection signatures.
3. Inter-MG Exchange: Reshaping tie-line power exchanges to disrupt quasi-static attacker models.
- Note: The action vector also includes a continuous islanding signal ( $\alpha_{iso}$ ) that triggers a binary isolation state via a hysteresis rule to prevent cascading failures.
Federated Architecture with Structural Privacy:
- Shared Encoder: A global encoder ( $\Phi$ ) is federated across all MGs to learn common representations of grid states.
- Private Heads: Each MG retains a private Actor and Critic head ( $\psi_i$ ). This design ensures that no MG exposes the cardinality, location, or configuration of its specific D-FACTS lines, BES units, or tie-line capacities, preserving structural privacy.
Byzantine-Resilient Aggregation:
- The server employs a two-stage robust aggregation rule.
- Stage 1: Coordinate-wise $\beta$ -trimmed-mean filtering to remove extreme outliers.
- Stage 2: A novel Reward-Weighted Aggregation coefficient ( $w_i$ ). This weight is calculated based on the client's episodic reward and a detection-quality score ( $\eta_{q,i}$ ), which combines the F1-score and False Positive Rate (FPR). Clients that generate excessive false alarms are automatically down-weighted, preventing them from poisoning the global model.
Differential Privacy (DP): To further protect privacy, Gaussian noise is added to encoder updates before transmission, clipped to an $\ell_2$ norm.
Adaptive Islanding: Upon detecting a critical compromise (via a multi-statistic voting detector), the local agent isolates the MG. A hysteresis mechanism enforces a minimum hold time ( $T_{iso,min}$ ) to prevent rapid switching oscillations, while proportional load shedding is applied to maintain local power balance.

3. Key Contributions

The paper identifies four primary contributions:

Byzantine-Resilient Framework: A unified framework integrating triple-surface MTD (reactance, storage, exchange) with adaptive islanding, specifically designed for decentralized MG clusters.
Federated Architecture with Privacy: A shared-encoder/private-head design that enables collaborative learning while maintaining the structural privacy of local defense device placements.
Reward-Weighted Aggregation Rule: A novel aggregation mechanism that penalizes clients causing false alarms or exhibiting compromised behavior, offering superior robustness against model poisoning compared to standard median or trimmed-mean filters.
Autonomous Adaptive Isolation: An intelligent islanding mechanism governed by a multi-statistic detector and hysteresis control, which isolates threats without relying on centralized commands.

4. Experimental Results

The framework was evaluated on two test systems: an interconnected cluster of four IEEE 30-bus MGs and a cluster of four IEEE 118-bus MGs.

Federated vs. Decentralized: The federated approach improved mean recall by 1.33 percentage points (118-bus) and 0.61 percentage points (30-bus) compared to decentralized baselines. Crucially, it demonstrated strong cross-MG generalization; when applied to an MG it had never trained on, the FL policy lost only 0.89–1.61 recall points, whereas decentralized policies suffered significantly higher degradation (1.17–5.36 points).
Byzantine Resilience: Under severe poisoning conditions (e.g., scaled Gaussian attacks with deviations 130x the baseline), standard aggregation methods (FedAvg, Krum) failed or suffered massive errors. The proposed Reward-Weighted rule achieved a 98.62% reduction in aggregation error.
False Positive Control: While standard robust filters (Trimmed Mean, Median) aggressively maximized True Positives, they inflated False Positive Rates (FPR). The Reward-Weighted rule significantly suppressed false alarms (e.g., 520 FPs vs. 914 for Median on the 118-bus system), prioritizing system stability.
Attack Mitigation: The framework achieved at least 95% burst recall against stealthy FDI attacks across all MGs. The MTD mechanism successfully forced pre-calculated attack vectors out of the valid state space.
Islanding Behavior: Adaptive islanding operated as a correlated control action rather than a reflex. In the 118-bus system, 77.8% of islanding events coincided with genuine attacks. The hysteresis mechanism successfully prevented rapid switching, and neighbor MGs were co-islanded only 20.3% of the time, preventing cascading failures.
Cost Impact: The framework reduced operational cost increments by up to 39.33% (30-bus) and 12.34% (118-bus) compared to non-federated baselines. The activation of defensive isolation incurred a negligible 1.66% increase in the aggregate operational costs of neighboring networks.

5. Significance and Claims

The paper claims that BR-FedMAPPO addresses a critical gap in cyber-physical security by unifying structural privacy, Byzantine resilience, and multi-surface physical defense within a single decentralized framework.

The authors assert that the framework is significant because:

It moves beyond reactive detection to proactive defense by dynamically altering the physical attack surface (Jacobian matrix) and control topology.
It solves the privacy paradox in multi-owner MGs, allowing collaborative learning without exposing sensitive topology or device configurations.
It introduces a behavioral feedback loop into federated aggregation, where the server dynamically scales client influence based on real-world detection quality, effectively neutralizing malicious actors without requiring complex cryptographic overhead.
It demonstrates that adaptive islanding can be learned as a cooperative control action that contains cyber-physical threats without causing catastrophic cascading failures, provided a hysteresis mechanism is employed.

The study concludes that while the framework increases islanding activity during genuine attacks (a necessary defensive expenditure), it maintains a high equilibrium between detection precision and operational stability, effectively neutralizing stealthy vulnerabilities that traditional static defenses cannot detect.