Adaptive Capsule Neural Network Framework Optimizes Edge Cloud Metadata Deduplication and Retrieval

A newly formulated computing framework integrating an Adaptive Capsule Neural Network with an improved optimization algorithm has demonstrated high-precision metadata deduplication, secure placement, and accelerated retrieval across distributed edge-cloud storage environments. Published in Discover Computing by researchers Radha V and R. Anitha, the investigation addresses the rapid expansion of metadata generated by smart industrial systems. In modern edge-cloud infrastructures, massive volumes of heterogeneous data gathered from distributed sources introduce severe redundancy, exhausting storage capacity and degrading analytical efficiency. To resolve these challenges, the authors introduced a framework named IBKA-A-CNet that couples machine learning duplicate elimination with optimized spatial placement and structured query extraction. The proposed framework achieved a metadata deduplication accuracy of 96.58 percent and an area under the receiver operating characteristic curve of 0.988. By concentrating reduction directly at the metadata tier rather than across raw payloads, the architecture lowers transmission latency, conserves network bandwidth, and preserves data integrity across decentralized computing environments.

At the core of the deduplication pipeline lies the Adaptive Capsule Neural Network, an architectural redesign of conventional capsule networks originally formulated for computer vision. Standard capsule networks utilize artificial neurons grouped into capsules to represent feature orientation, size, and position, employing dynamic routing to preserve hierarchical feature relationships. However, conventional designs struggle when applied to sparse, non-visual, and heterogeneous metadata records, encountering vanishing gradient complexities and training instability under massive data volumes. To address these operational constraints, the authors implemented a specialized metadata embedding layer at the input stage. This initial embedding converts heterogeneous metadata attributes into dense feature vectors, allowing the network to capture semantic correlations before feature extraction occurs. Furthermore, the architecture introduces stacked convolutional layers that progressively extract hierarchical metadata representations, followed by a dedicated metadata capsule layer utilizing dynamic routing, a fully connected layer, a dropout layer to reduce overfitting, and a terminal sigmoid classification function.

To automate the tuning of the neural network and avoid arbitrary manual configurations, the researchers formulated the Improved Black-Winged Kite Algorithm, an evolutionary heuristic modeled on the migratory and predatory behaviors of the black-winged kite. In natural settings, these raptors adjust flight velocity according to wind conditions when tracking prey before initiating sharp descending attacks. The standard optimization algorithm translates these behaviors into mathematical exploration and exploitation operators, integrating leader bird evaluation and a Cauchy mutation distribution to navigate complex solution spaces. Nevertheless, standard implementations frequently suffer from premature convergence and computational bottlenecks during fitness evaluation phases due to fixed random search coefficients. To overcome these limitations, the authors redesigned the algorithmic formulation by replacing static random variables with a population-fitness-dependent adaptive parameter. This adaptive factor dynamically evaluates candidate solutions relative to mean population fitness, effectively shifting the search balance toward broader global exploration when population diversity is high and refining local exploitation as candidate solutions converge.

Within the deduplication architecture, the Improved Black-Winged Kite Algorithm optimizes three critical hyperparameters of the Adaptive Capsule Neural Network: activation function choice, training epoch count, and hidden neuron allocations. The research team defined an objective function targeting the inverse of classification precision, directing the heuristic search to minimize false positive identifications while maintaining reliable convergence. The optimization process operates within bounded search domains, evaluating activation functions across five index settings, constraining epoch counts between 5 and 50 iterations, and searching hidden neuron allocations across an integer range spanning from 5 to 255. Bounding these parameters prevents excessive network complexity, which would otherwise induce overfitting and training delays, while avoiding overly restricted configurations that cause gradient disappearance. By executing iterative fitness evaluations across candidate solutions, the optimization scheme fine-tunes network capacity, allowing it to adapt to diverse metadata distributions and maintain high discrimination accuracy without numerical instability.

Beyond identifying duplicate records, the framework coordinates the physical distribution of metadata across geographically dispersed edge devices using an objective placement strategy driven by the optimization algorithm. Traditional metadata placement methodologies relying on standard clustering techniques often suffer from load imbalances, elevated transmission delays, and high coupling across distinct storage facilities. The proposed methodology resolves these spatial challenges by formulating an objective function that minimizes the Euclidean distance between candidate edge server nodes and the virtual coordinates of incoming data within the spatial domain. Because edge computing topologies are inherently dynamic, with nodes frequently joining, disconnecting, or altering operational states, the optimization algorithm continuously recalculates Euclidean distances across time-varying network intervals. When substantial topological shifts occur, the algorithm increases exploratory search behavior while simultaneously leveraging previously identified high-quality deployment coordinates, avoiding complete reinitialization and ensuring metadata is positioned on edge nodes that are both geographically proximate and computationally available.

To ensure comprehensive data confidentiality and mitigate vulnerabilities associated with distributed network transmission, all metadata undergoes cryptographic protection prior to physical placement using Elliptic Curve Cryptography. In contrast to conventional cryptosystems such as the Advanced Encryption Standard in counter mode or Rivest-Shamir-Adleman algorithms, elliptic curve techniques provide equivalent security guarantees with shorter key lengths and lower memory requirements. The implemented scheme computes mathematical points along an elliptic curve defined by cubic equations, generating private and public key pairs that protect sensitive metadata records from unauthorized manipulation during transit and storage. Security evaluations assessing resistance against known-plaintext attacks and chosen-plaintext attacks demonstrated that the encryption mechanism maintains strong cryptographic protection across varying operational conditions. In experimental attack scenarios measuring statistical correlation coefficients between original plaintext records and adversary-reconstructed data, the proposed framework achieved a correlation coefficient of 0.35 percent, outperforming the 0.70 percent coefficient recorded by baseline algorithms and preventing unauthorized key reconstruction.

Once encrypted metadata is deposited across the edge-cloud topology, access and extraction are managed through an adapted Data Retrieval Protocol designed to execute queries without introducing latency bottlenecks or compromising encryption boundaries. The retrieval pipeline is structured across five functional tiers comprising chunking, hashing, indexing, mapping, and storage layers. In this operational sequence, user requests are initially compiled into formal query word vectors, followed by the generation of cryptographic trapdoor values computed using shared secret keys established between data owners and authorized clients. The edge-cloud service provider verifies these trapdoors against stored index structures and returns extended tuple representations. Upon decrypting intermediate server responses, the client expands the query scope, executes a secondary trapdoor verification to receive exact document identifiers, and applies a ranking function to reconstruct and decrypt targeted metadata files using the private key, maintaining data integrity and high availability across high-volume querying workloads.

Empirical validation of the framework was executed in a Python computational environment using the Kaggle Dataset Search Metadata collection, which aggregates large volumes of structured and semi-structured metadata records extracted from distributed cloud repositories. Deduplication performance was evaluated across five-fold cross-validation trials, with the fifth fold demonstrating the highest consistency across evaluation metrics. In epoch-based comparative analyses against baseline deduplication systems including blockchain-assisted schemes, modified elliptic curve models, two-threshold divisor chunking, and standard capsule networks, the optimized capsule network attained an accuracy of 94.41 percent. When evaluating the optimization algorithm independently, the Improved Black-Winged Kite Algorithm achieved 97.25 percent accuracy at 500 training epochs. Statistical evaluations involving five independent executions revealed that the proposed model achieved the lowest objective function value of 0.772402, a mean value of 0.864318, a median of 0.851924, and a standard deviation of 0.142836, outperforming alternative optimization techniques including wild geese migration optimization and mayfly optimization algorithms.

Dedicated evaluations of the spatial placement and retrieval subsystems further confirmed the operational viability of the integrated framework. In standalone placement benchmarks, the optimized placement mechanism attained an allocation accuracy of 97.63 percent, achieving a placement latency of 71 milliseconds, a placement overhead of 16.47 percent, a response time of 79 milliseconds, and a load balancing efficiency of 94.18 percent. Detailed cumulative overhead assessments revealed that metadata placement consumed approximately 12.4 milliseconds of latency, 7 kilobytes of network overhead, 18 percent processor utilization, and 56 megabytes of memory. Concurrently, the Data Retrieval Protocol recorded a query execution duration of 482 milliseconds, providing reliable retrieval precision while reducing communication delays. In end-to-end testing, the architecture supported a sustained throughput of 936 requests per second, which increased to 951 requests per second under low-load conditions and demonstrated scalable throughput reaching 980 percent at 10,000 concurrent requests, confirming minimal computational overhead.

Despite favorable experimental outcomes, the authors highlighted several technical limitations inherent to the current implementation that define focal points for future research. Specifically, the framework exhibits increased computational complexity when processing exceptionally high-dimensional training datasets, introducing execution overhead and potential resource constraints under extreme operational bursts. Furthermore, the architecture currently lacks thermal-aware resource management modules capable of monitoring hardware temperature profiles during continuous duplicate identification across dense edge clusters. The study also noted the absence of preliminary data preprocessing and independent feature extraction stages prior to neural network ingestion, which influences input data quality under noisy conditions. To resolve these challenges, future investigations will explore the integration of lightweight federated learning frameworks to support decentralized training across edge nodes without centralizing raw metadata, alongside automated preprocessing pipelines and alternative cryptographic retrieval mechanisms designed to maintain energy efficiency and long-term sustainability in expansive industrial computing networks.

Subject of Research: Machine learning framework for metadata deduplication, placement, and retrieval in edge-cloud computing

Article Title: A metadata deduplication framework associated with metadata placement and retrieval in edge-cloud using Adaptive Capsule Neural Network (A-CNet)

Article References: V, R., & Anitha, R. (2026). A metadata deduplication framework associated with metadata placement and retrieval in edge-cloud using Adaptive Capsule Neural Network (A-CNet). Discover Computing, 29(1), Article 766. https://doi.org/10.1007/s10791-026-10680-2

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10680-2

Keywords: metadata deduplication, edge-cloud computing, Adaptive Capsule Neural Network, Improved Black-Winged Kite Algorithm, Elliptic Curve Cryptography, Data Retrieval Protocol, metadata placement, Euclidean distance optimization, CapsNet architecture, data redundancy reduction, known-plaintext attack resistance, chosen-plaintext attack resistance, hyperparameter optimization, distributed storage systems, query execution latency, load balancing efficiency, receiver operating characteristic