A research team led by Swathi Tejah Yalla, Isha Batra, and Arun Malik has introduced a specialized deep learning framework designed to detect denial-of-wallet attacks within serverless cloud computing environments. In simulated benchmark evaluations, the proposed architecture achieved an overall testing detection accuracy of 95.30 percent alongside a training accuracy of 95.62 percent. Serverless computing and Function-as-a-Service architectures automate backend resource allocation and rely on pay-per-use billing models, which provide substantial operational flexibility for developers. However, these identical architectural characteristics create novel threat surfaces where malicious actors can exploit auto-scaling mechanisms to generate prolonged, illicit resource consumption. Rather than attempting to disable infrastructure, such attacks target the economic layer by artificially inflating service provider invoices without necessarily producing observable performance degradation.
Unlike traditional denial-of-service incursions that overwhelm operational throughput, denial-of-wallet threats typically manifest through seemingly legitimate function invocations. This operational profile makes surface-level differentiation via request volume or latency metrics extremely difficult. The authors noted that when automated scaling, dynamic execution, and pay-for-use billing converge, even minor patterns of persistent misuse can generate catastrophic financial losses for organizations deploying digital wallet services. In response to these vulnerabilities, the researchers formulated the Gated Recurrent Unit Networks Digital Wallet Attack Detection framework, designated as GRU-DWAD or GRN-DWAD, as an integral component of a broader Cloud Security Assurance Framework. The primary objective centered on establishing a precise, interpretable, and computationally scalable defense mechanism capable of tracking sequential dependencies and autocorrelation patterns across execution data streams.
The methodological pipeline of the framework operates across three distinct computational stages comprising data pre-processing, wrapper-based feature selection, and recurrent sequence classification. In the initial stage, input metrics undergo statistical normalization to achieve uniform scaling across dynamic serverless variables. The documentation notes that Z-score normalization and min-max normalization were employed to prevent numerical distortions across varying evaluation scales, ensuring that dynamic execution characteristics do not compromise model convergence during parameter training. By establishing consistent input distributions, the normalization routine mitigates erratic weight updates and assists the downstream classifier in isolating genuine anomalies in resource consumption patterns that point toward cost-driven exploitation.
To reduce computational overhead while preserving diagnostic capability, the authors integrated Recursive Feature Elimination using a decision tree estimator, designated as DT-RFE. Starting from an initial baseline of 18 recorded attributes, the backward elimination search iteratively evaluated and discarded redundant variables that failed to improve predictive accuracy. The procedure incorporated 10-fold cross-validation through StratifiedKFold splitting to preserve proportional representation across classes during every evaluation cycle. Through this recursive pruning routine, the original feature space was refined down to 11 critical attributes deemed essential for identifying financial exploitation: client IP address, submit time, invocation delay, function trigger, round-trip time, timestamp, active functions at request, response delay, function duration, active functions at response, and maximum CPU utilization.
The selected attributes reflect an integrated view of both temporal behavior and infrastructure resource consumption, capturing the subtle micro-delays and concurrency footprints left by automated attack scripts. Following feature selection, the refined temporal sequences enter a fine-tuned recurrent neural network architecture based on Stacked Long Short-Term Memory units, supplemented by Gated Recurrent Unit principles. To counter overfitting across dense representations, the network incorporates L2 normalization on internal parameters alongside a dropout layer configured with a 0.5 dropout rate positioned after the initial fully connected layer. The sequential memory cell operates through three functional gates: a forget gate that discards outdated information from preceding states, an input gate that incorporates candidate values derived from current inputs and prior hidden states, and an output gate that calculates final hidden representations using hyperbolic tangent and sigmoid activations.
Experimental evaluation was executed using Python 3.10 and open-source machine learning libraries including TensorFlow, Keras, Scikit-learn, Pandas, NumPy, Matplotlib, and Seaborn. The validation utilized a benchmark dataset comprising 10,000 instances equally divided between 5,000 legitimate transactions and 5,000 attack events to avoid class imbalance biases. The evaluation data were synthesized via a hybrid configuration that combined logs from the Digital Wallet Attack Test Simulator, known as DWATS, with synthetically generated attack traces designed to simulate invocation delays, billing discrepancies, and resource exhaustion behaviors typical of cloud function environments. Training was conducted using the Adam optimization algorithm configured with a learning rate of 0.001, a batch size of 64, and a maximum ceiling of 25 training epochs.
The resulting learning trajectories revealed consistent optimization dynamics over the course of training, with both training and validation accuracy steadily ascending before reaching convergence around the 20th epoch with minimal variance between curves. In performance assessments across an 80:20 training-to-testing split representing 8,000 training and 2,000 testing instances, the model maintained high classification fidelity across both legitimate and malicious traffic. Analysis of the confusion matrix demonstrated misclassification rates remaining below 7 percent for both categories. Across the complete test cohort, the framework attained a precision metric of 94.85 percent and a recall metric of 94.12 percent, demonstrating balanced sensitivity and specificity under controlled simulation settings.
Comparative evaluations against standard baseline architectures highlighted the advantages of combining recurrent sequence modeling with wrapper-based feature reduction. When benchmarked against alternative intrusion detection paradigms, the proposed recurrent model outperformed Convolutional Neural Networks, standard Long Short-Term Memory baselines, Gated Recurrent Units alone, Random Forest classifiers, Support Vector Machines, and Autoencoder architectures. The experimental findings indicated a 3 to 6 percent improvement in detection accuracy over standalone LSTM and GRU baselines, alongside an advantage of 7 to 10 percent over classical machine learning classifiers such as Random Forest and Support Vector Machines, confirming the utility of capturing temporal autocorrelation in serverless audit logs.
The authors contextualized their findings within the broader trajectory of intrusion detection literature, noting that previous research into distributed learning, federated optimization, and blockchain verification has addressed traditional service disruption but frequently overlooked subtle economic denial mechanisms. Prior frameworks such as Gringotts, PrivFlow, Valve, and DWATS highlighted the need for function-level and workflow-level observability to prevent fiscal exhaustion. By demonstrating that recurrent networks can identify coordinated resource-draining behaviors from execution logs, the study provides empirical evidence supporting the integration of automated sequence modeling into operational cloud monitoring systems.
Despite the high detection metrics reported in the study, the authors documented several technical constraints that bound the generalizability of their findings. The empirical validation relied exclusively on a simulated benchmark dataset derived from hybrid DWATS logs, which may not capture the full heterogeneity, unpredictable latency shifts, or novel attack variations present in production multi-tenant cloud ecosystems. Additionally, the architecture was evaluated within a uniform testing setup rather than across decentralized or multi-cloud deployments where distributed execution traces could introduce asynchronous communication overhead. Future research directions outlined by the authors focus on extending validation across heterogeneous and hybrid cloud platforms, developing adaptive learning mechanisms to track evolving attack vectors, and optimizing real-time processing efficiency to support continuous threat screening.
The operational logic behind denial-of-wallet incursions reflects a deliberate departure from conventional network disruptions. In typical service attacks, adversaries seek to overwhelm system bandwidth or crash server hosts to prevent legitimate client access. Within serverless Function-as-a-Service environments, however, applications scale automatically to accommodate incoming requests, dynamically provisioning virtual resources and invoicing operators on a strict pay-as-you-go schedule. Attackers exploit these elastic billing mechanics by dispatching streams of seemingly genuine invocations that stay below standard alerting thresholds. Because server response latency and total throughput can appear entirely normal, these micro-scale abuses generate no acute operational failures while silently driving computing charges to ruinous levels.
To contextualize their defense strategy, the researchers examined prior attempts to safeguard dynamic distributed ecosystems against related threat vectors. Earlier investigations evaluated methods such as grey wolf optimization to counter grey hole and black hole routing exploits in mobile ad hoc networks, demonstrating that heuristic optimization can trace subtle behavioral anomalies. Parallel lines of research explored neutrosophic hypersoft sets to capture mathematical uncertainty in incomplete attack signatures, while clustering algorithms reduced computational overhead in dynamic routing management. Similarly, specialized platforms like Valve demonstrated the necessity of inspecting serverless execution workflows to limit vulnerability surfaces, reinforcing the premise that financial defense requires continuous monitoring across both individual function triggers and broader pipeline orchestrations.
Broader intrusion literature has also assessed decentralized distributed architectures, including federated learning pipelines integrated with blockchain ledgers, to train collaborative threat detectors without exposing sensitive local data. Studies examined asynchronous federated frameworks combining bidirectional recurrent networks with attention layers to capture low-rate distributed denial-of-service threats across heterogeneous devices. Concurrently, explainability frameworks employing Shapley Additive Explanations have emerged to demystify complex neural network classifications in industrial environments where stakeholders require transparent rationales for flagged anomalies. The authors emphasized that while distributed federated paradigms advance collective privacy, specialized deep sequential modeling remains essential for parsing the nuanced temporal dynamics unique to cloud billing records.
Subject of Research: Deep learning detection of denial-of-wallet cyberattacks in serverless computing architectures
Article Title: Gated Recurrent Unit-Based Detection Framework for Digital Wallet Attacks in Serverless Computing Environments
Article References: Yalla, S. T., Batra, I., & Malik, A. (2026). Gated Recurrent Unit-Based Detection Framework for Digital Wallet Attacks in Serverless Computing Environments. International Journal of Networked and Distributed Computing, 14(1), Article 10. https://doi.org/10.1007/s44227-025-00084-x
Image Credits: AI Generated
DOI: 10.1007/s44227-025-00084-x
Keywords: serverless computing, cloud security, denial of wallet attack, digital wallet security, function as a service, gated recurrent units, stacked lstm, recursive feature elimination, anomaly detection, cybersecurity, machine learning, deep learning, cloud billing exploitation, feature selection, intrusion detection systems, dwats simulator
Cite Scienmag News
Blake Davidson. (November 20, 2026). Recurrent Neural Network Framework Detects Serverless Digital Wallet Attacks With High Accuracy. Scienmag. https://scienmag.com/recurrent-neural-network-framework-detects-serverless-digital-wallet-attacks-with-high-accuracy/
Blake Davidson. "Recurrent Neural Network Framework Detects Serverless Digital Wallet Attacks With High Accuracy." Scienmag, 20 November 2026, https://scienmag.com/recurrent-neural-network-framework-detects-serverless-digital-wallet-attacks-with-high-accuracy/. Accessed 20 November 2026.
Blake Davidson. "Recurrent Neural Network Framework Detects Serverless Digital Wallet Attacks With High Accuracy." Scienmag. November 20, 2026. https://scienmag.com/recurrent-neural-network-framework-detects-serverless-digital-wallet-attacks-with-high-accuracy/








