Federated Learning and Multi-Objective Optimization Improve Mobile Ad Hoc Network Security

Researchers Diksha Shukla and Raghuraj Singh have formulated a decentralized cybersecurity framework that significantly improves data delivery and route defense across mobile ad hoc networks vulnerable to disruptive packet-dropping intrusions. Mobile ad hoc networks, commonly abbreviated as MANETs, operate without fixed physical infrastructure or centralized administrative hubs, relying instead on autonomous mobile nodes that dynamically establish wireless links to relay communication traffic across shifting operational environments. Because these decentralized topologies depend entirely on mutual cooperation among independent mobile nodes to forward intermediate packets, they remain highly susceptible to malicious intrusions such as black hole attacks and gray hole attacks, where compromised or rogue devices falsely advertise optimal routing paths and subsequently discard data transmissions. The newly designed system integrates long short-term memory recurrent neural networks for continuous dynamic trust evaluation, federated learning for privacy-preserving collaborative model updates across distributed nodes, and the non-dominated sorting genetic algorithm II to balance conflicting routing priorities including trustworthiness, latency, and energy efficiency.

The fundamental operational architecture of conventional MANET routing protocols, such as the widely deployed Ad hoc On-Demand Distance Vector protocol, relies on trust metrics to identify reliable transmission paths, yet conventional defense mechanisms frequently falter under dynamic network conditions. Traditional trust estimation techniques typically depend on static thresholds, rule-based counters, or centralized management authorities that assess direct and indirect packet forwarding ratios or acknowledgment receipts. However, static rules fail to adapt quickly when mobile nodes alter their behaviors or move through volatile topologies, while centralized supervisory nodes introduce severe scalability bottlenecks, single points of failure, and unacceptable computational latency across bandwidth-constrained wireless channels. Furthermore, conventional machine learning techniques that analyze behavioral anomalies demand substantial volumes of centralized historical training data, an operational requirement that directly conflicts with the decentralized nature of ad hoc networks and introduces severe data privacy vulnerabilities among participating mobile devices.

To overcome these structural limitations, the research team structured a multi-tiered framework designated as FLT-AODV that distributes behavioral modeling directly to the network periphery while coordinating collaborative learning without centralizing raw traffic data. At the local level, each individual node hosts an internal long short-term memory neural network configured to ingest a sequential feature vector capturing twelve distinct static and dynamic interaction metrics over time. Rather than judging a neighboring node by instantaneous or threshold-based performance snapshots, the recurrent architecture evaluates temporal behavioral trends across successive routing requests, reply exchanges, packet forwarding consistencies, latency variations, and drop frequencies. This continuous sequential evaluation enables the local model to generate a refined mathematical trust score ranging from zero to one, where lower numerical values systematically isolate malicious nodes exhibiting the distinct packet-dropping profiles characteristic of black hole and gray hole adversaries.

Collaborative intelligence is achieved across the moving network using a decentralized federated learning scheme governed by node encounter probabilities derived from mobility patterns. Under this mechanism, individual mobile nodes do not transmit raw packet logs, private traffic histories, or detailed communication records across the shared wireless channel, thereby preserving strict data privacy and sharply curtailing transmission overhead. Instead, whenever a dedicated federated learning timer triggers, neighboring nodes compute and exchange localized model weight updates. Each receiving node then incorporates these updates using a weighted aggregation formula scaled by encounter probabilities and a predefined learning rate, dynamically refining its local predictive neural network. This cooperative aggregation enables individual nodes to rapidly assimilate emerging behavioral anomalies observed elsewhere in the topology, producing robust, sequence-aware threat detection that scales seamlessly despite rapid mobility and continuous topological reconfiguration.

Recognizing that maximizing security in isolation can degrade standard quality of service parameters, the framework directs the updated trust scores into a multi-objective optimization layer governed by the non-dominated sorting genetic algorithm II, or NSGA-II. In mobile routing environments, network objectives are inherently conflicting: selecting nodes solely for peak trustworthiness can lengthen routing paths, increase end-to-end transmission delays, and rapidly exhaust the battery reserves of specific trusted devices. To resolve these operational trade-offs, the optimization algorithm treats route selection as a multi-objective problem across a continuous decision variable representing an objective weight preference ratio. The genetic algorithm systematically executes binary tournament selection based on Pareto rank and crowding distance, followed by simulated crossover and mutation operations across successive generations, to derive a Pareto-optimal front of candidate routing options balancing trustworthiness, residual energy, and transmission latency.

The operational scheduling of the framework is coordinated through four synchronized background timers that govern feature extraction, local neural network training, federated weight sharing, and routing information updates. The feature extraction module operates across fixed observation windows aligned with standard Ad hoc On-Demand Distance Vector control packets, while the long short-term memory training routine accumulates predefined sequence lengths to maintain temporal learning stability. Crucially, the federated model aggregation rounds run asynchronously, accommodating intermittent wireless disconnections and variable node velocities without enforcing rigid global synchronization across the entire ad hoc cluster. Once the Pareto front generates the optimal weighting vector for current operational constraints, the underlying routing engine dynamically chooses the most resilient next-hop nodes, progressively excluding low-trust actors from active communication paths.

To rigorously evaluate the system, the authors conducted extensive simulation experiments using the Network Simulator 2 platform, specifically version 2.35, across a standardized two-thousand-meter by two-thousand-meter virtual landscape. Physical and medium access control operations adhered to the IEEE 802.11 wireless standard using omnidirectional antennas and a two-ray ground propagation model, while node mobility was simulated via the Random Waypoint Model. Network configurations ranged from 100 to 2000 active mobile nodes initialized with 100 Joules of energy, with data traffic configured as constant bit rate streams transmitting 512-byte packets over UDP for a duration of 600 seconds. To simulate severe adversarial pressure, exactly ten percent of the active nodes were designated as malicious, partitioned into black hole nodes that dropped all received packets after falsely claiming optimal routes, and gray hole nodes that selectively discarded data to evade detection.

The resulting performance data demonstrated substantial operational gains across all primary evaluation metrics when compared against standard routing baselines and competing hybrid architectures. In baseline Ad hoc On-Demand Distance Vector networks subjected to adversarial attacks, packet delivery ratios collapsed to 44 percent in 100-node networks and reached only 62 percent at 2000 nodes due to unchecked packet dropping. In contrast, the integrated federated framework maintained a packet delivery ratio exceeding 98.6 percent even at the lowest node density of 100 nodes, ultimately reaching a peak delivery ratio of 99.9 percent in dense 2000-node scenarios. This performance markedly surpassed existing benchmark protocols from prior literature, including cluster-based artificial bee colony optimization combined with mutation-based artificial neural networks, which achieved an upper delivery ceiling of 99.08 percent at 1000 nodes.

Throughput and latency measurements exhibited similar resilience under dense traffic conditions and severe hostile interference. Standard routing without security protections saw its throughput degrade severely under attack, plummeting from an unattacked range of 82.67 to 91.6 kilobits per second down to between 58.97 and 67.85 kilobits per second. The proposed framework reversed these losses entirely, sustaining throughput between 98.92 kilobits per second at lower densities and 99.8 kilobits per second across 2000 nodes, outperforming comparative state-of-the-art benchmarks that peaked at 97.86 kilobits per second. Simultaneously, the framework reduced end-to-end transmission latency to 0.015 seconds for 100 nodes and 0.068 seconds for 2000 nodes, achieving an overall minimum recorded delay of 0.035 seconds compared to 0.045 seconds and 0.071 seconds reported by prior comparative methodologies.

To confirm that the recorded improvements reflected genuine protocol efficacy rather than stochastic artifacts of simulation modeling, the authors repeated all experimental scenarios across ten independent runs with varying random number seeds. Statistical validation confirmed a mean packet delivery ratio of 98.94 percent with a standard deviation of 0.12 and a 95 percent confidence interval of plus or minus 0.10, alongside a mean throughput of 97.63 kilobits per second and a mean end-to-end delay of 0.046 seconds. Student t-test evaluations comparing the proposed model against existing methods yielded t-values of 4.82 for packet delivery, 5.14 for throughput, and 4.37 for transmission delay, all achieving p-values below 0.05 to confirm that the observed performance advantages are statistically significant.

Despite its demonstrated effectiveness, the authors noted specific engineering limitations inherent to combining multi-objective evolutionary computation with recurrent deep learning. The local execution of recurrent neural layers, frequent feature logging, and parameter exchanges introduce measurable computational overhead and communication costs, which may constrain execution on severely energy-restricted, lightweight hardware nodes. Furthermore, the current evaluation examined only black hole and gray hole attacks within simulated random waypoint mobility models, leaving advanced threat vectors such as Sybil, wormhole, and Byzantine attacks unaddressed. The authors plan to investigate lightweight optimization models and broader adversarial scenarios in subsequent research.

Subject of Research: Intrusion detection and multi-objective routing optimization in mobile ad hoc networks

Article Title: Multi-objective optimization of federated intrusion detection system for detecting black hole and gray hole attacks in MANET

Article References: Shukla, D., & Singh, R. (2026). Multi-objective optimization of federated intrusion detection system for detecting black hole and gray hole attacks in MANET. Discover Electronics, 3(1), Article 108. https://doi.org/10.1007/s44291-026-00258-z

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00258-z

Keywords: mobile ad hoc networks, MANET, intrusion detection system, federated learning, NSGA-II, LSTM, AODV protocol, black hole attack, gray hole attack, multi-objective optimization, network security, packet delivery ratio, end-to-end delay, network throughput, dynamic trust prediction, NS-2 simulation, decentralized learning