Denial-of-Wallet (DoW) Attacks
- Denial-of-Wallet (DoW) is an attack vector that exploits pay-per-use billing to drain financial resources without necessarily disrupting service availability.
- The attack leverages serverless function invocations and blockchain transaction dynamics through low-rate or amplified request patterns to induce cost exhaustion.
- Mitigation strategies focus on cost-aware throttling, anomaly detection in spend rates, and adaptive billing safeguards to preserve economic sustainability.
Searching arXiv for recent and foundational papers on Denial-of-Wallet, especially serverless/FaaS and broader taxonomy. Denial-of-Wallet (DoW) denotes a class of denial attacks whose primary objective is financial exhaustion rather than immediate service unavailability. In serverless computing, the foundational formulation defines DoW as “the intentional mass, and continual, invocation of serverless functions, resulting in financial exhaustion of the victim in the form of inflated usage bills” (Kelly et al., 2021). Subsequent work places DoW within the sustainability branch of denial attacks, distinguishing it from availability-centric DoS and DDoS by its exploitation of elastic auto-scaling, fine-grained metering, and serverless billing semantics (Dorsett et al., 24 Aug 2025). In adjacent literatures, the same term or a closely related mapping is also used for wallet-specific unavailability in blockchain systems and cryptocurrency wallets, where the result is an inability to sign, access funds, or obtain transaction inclusion rather than cloud-billing exhaustion (Urien, 2023).
1. Definition and conceptual scope
The canonical serverless definition is the one introduced in “Denial of Wallet -- Defining a Looming Threat to Serverless Computing” (Kelly et al., 2021): DoW attacks are “the intentional mass, and continual, invocation of serverless functions, resulting in financial exhaustion of the victim in the form of inflated usage bills.” This formulation anchors DoW in event-driven, pay-as-you-go environments where invocations, execution duration, and associated service activity are billable.
A later taxonomy formalizes DoW as an extension of Economic Denial of Sustainability targeting serverless Function-as-a-Service infrastructure. In that framing, DoW “operate[s] with minimal resource impact but exploit[s] billing granularity,” and therefore differs from classical denial attacks that attempt to saturate network or compute capacity (Dorsett et al., 24 Aug 2025). The central distinction is not whether requests are malicious in the abstract, but whether their operational effect is to drain budgets while leaving the service apparently healthy.
A comprehensive review of serverless DoW research reinforces this distinction by characterizing DoW as an economically motivated denial attack that exploits pay-as-you-go billing “without necessarily degrading service availability” (Dorsett et al., 24 Aug 2025). This suggests that DoW is best understood as a sustainability threat: the system may continue to answer requests, satisfy latency objectives, or auto-scale correctly, yet still fail operationally because the economic substrate becomes untenable.
Outside serverless computing, the term is used more loosely. In blockchain wallet security, protective lockout states that refuse signing or wallet operation when integrity and authenticity cannot be established can be interpreted as a wallet-specific form of DoW, although the paper in question does not itself use the term (Urien, 2023). In MimbleWimble, “transaction denial of service with aggregations” is explicitly mapped to a Denial-of-Wallet outcome because valid transactions fail to reach confirmation, denying the wallet’s ability to complete a payment (Tabatabaee et al., 2021). These usages share the endpoint of denied wallet function, but they are conceptually distinct from the serverless cost-exhaustion lineage.
2. Serverless mechanics that make DoW possible
Serverless environments combine event-driven execution, scale-to-zero, elastic auto-scaling, and usage-based billing. The 2021 definition paper describes serverless computing as a framework for “event driven, pay-as-you-go functions in a highly scalable environment,” and identifies those same properties as the reason a new attack paradigm emerges (Kelly et al., 2021). In practical terms, the attack surface is created by the linear or amplified translation of events into spend.
The serverless anomaly-detection vision paper explains why this translation is both powerful and difficult to observe. Functions are billed per invocation and per compute-time, often with additional charges for network egress and downstream services. Executions are short-lived and isolated, monitoring granularity is inconsistent, and observability across distributed triggers and dependent services is limited. As a result, DoW often appears as a “silent failure”: the application continues to work while budgets are drained (Nguyen et al., 7 Jul 2025). The paper further notes that 85% of tasks complete within one minute, which complicates the construction of stable baselines for anomaly detection in ephemeral, stateless systems.
The unified denial taxonomy identifies two conditions that separate DoW from older forms of denial: malicious requests target scalable cloud infrastructure resources, and they target serverless infrastructure specifically (Dorsett et al., 24 Aug 2025). In its notation, DoW is classified by the conjunction of maliciousness, a single source, cloud elasticity, and serverless targeting. This formalization is important because it makes explicit that DoW is not defined by raw traffic volume. A low-rate request stream can still qualify if it exploits elastic metering and serverless invocation paths.
The review literature emphasizes the same underlying exposure. Serverless platforms bill the “number of invocations” and “execution duration,” and they expose triggers such as HTTP endpoints, storage events, and queues. Combined with limited infrastructure visibility and abstraction, this enables attacks in which even “successful” or rejected events escalate cost without obvious service failure (Dorsett et al., 24 Aug 2025). A plausible implication is that serverless resilience mechanisms designed to preserve availability can, under DoW conditions, preserve functionality while worsening the financial outcome.
3. Attack patterns and amplification paths
Derived attack patterns described alongside the 2021 paper include HTTP or API floods to serverless endpoints, storage-trigger floods, queue or message-trigger floods, chained function amplification, egress-cost amplification, third-party webhook abuse, and replay or infinite-loop misconfiguration (Kelly et al., 2021). Although those patterns are presented as derived models consistent with the paper’s narrative rather than verbatim experimental taxonomies, they align with the attack surface implied by serverless trigger design.
The 2025 review organizes DoW into three primary types: Blast DDoW, Continual Inconspicuous DDoW, and Background Chained DDoW (Dorsett et al., 24 Aug 2025). Blast DDoW uses short, high-intensity invocation spikes to force rapid auto-scaling and sudden cost surges. Continual Inconspicuous DDoW emphasizes long-term, low-rate cost drain that blends with background traffic. Background Chained DDoW exploits orchestration and inter-function dependencies so that a single event triggers cascades of downstream invocations. These categories differ mainly in traffic shape, stealthiness, and the degree to which inter-function amplification is exploited.
The anomaly-detection vision paper presents closely related vectors: event storming via public endpoints, trigger abuse of storage and queues, recursive or fan-out orchestration, misconfigurations causing loops, cold start amplification, and cross-function cascades (Nguyen et al., 7 Jul 2025). Its case study is nominally about DoS, but the reported dynamics are directly relevant to DoW: short bursts of 59 req/s for 3.9 s and 203 req/s for 4.9 s caused prolonged recovery, response time above 15–20 s persisting beyond the burst, and function instances peaking at 40. The paper argues that benign-looking traffic can trigger cold-start thrashing and queuing delays that inflate compute-time and therefore cost.
The review also highlights concrete cost amplification via non-compute services. It cites an S3 case where rejected requests still incurred substantial daily cost, with a documented figure of up to USD $1,300/day (Dorsett et al., 24 Aug 2025). This is important because it shows that the billable surface is not limited to successful application logic. A plausible implication is that cost amplification can occur at multiple layers simultaneously: invocation billing, storage-triggered functions, downstream APIs, and transfer charges.
Function chaining introduces multiplicative growth. The review gives the parametric expression
$N_{\text{chain}}=\frac{f^{L+1}-1}{f-1},f=10L=31111Nc_{req}c_{GBs}M_it_iC = c_{req} N + c_{GBs}\sum_{i=1}^{N} M_i t_i.f=10$0 over horizon $f=10$1, the expected cost becomes
$f=10$2
The same material defines instantaneous spend rate and time to exhaust a budget $f=10$3 by
$f=10$4
These are explicitly labeled derived models rather than quoted equations from the paper (Kelly et al., 2021).
The review paper offers a symbolic total-cost decomposition consistent with invocations, execution duration, data transfer, and storage events:
$f=10$5
and an expected-cost expression under arrival rate $f=10$6:
$f=10$7
Because the review does not provide exact provider prices, it keeps the model symbolic except where a real-world monetary figure is available (Dorsett et al., 24 Aug 2025).
The unified taxonomy contributes a different formalism: classification by observable conditions $f=10$8–$f=10$9 (Dorsett et al., 24 Aug 2025). For DoW, the necessary and sufficient conjunction is
$L=3$0
where malicious requests exist, they come from a single source, they target scalable cloud infrastructure resources, and they target serverless infrastructure. In the paper’s lattice model, DoW occupies the vector
$L=3$1
This shifts analysis from naming patterns to evaluating observed predicates in telemetry and billing systems.
The same taxonomy also formalizes low-rate behavior through condition $L=3$2:
$L=3$3
DoW does not require $L=3$4, but the framework makes clear that stealth and low-rate properties can coexist with financial exhaustion (Dorsett et al., 24 Aug 2025). This suggests a useful separation between attack objective and traffic morphology: DoW is defined by what is being exhausted, not solely by how traffic is shaped.
5. Detection and empirical study
The 2021 paper reports an “Isolated Test Platform” and notes that OpenFaaS no longer supported Docker Swarm, leading to a migration to Kubernetes with experiments re-run (Kelly et al., 2021). The available excerpt does not include experimental parameters, figures, or quantitative results. Its importance lies mainly in establishing safe research infrastructure for studying DoW without incurring production costs.
Later work expands the empirical and detection landscape. The serverless anomaly-detection vision paper argues that threshold-based methods developed for long-running, stateful services are ill-suited to serverless because of absent persistent state, inconsistent monitoring granularity, and the difficulty of correlating behavior across isolated functions (Nguyen et al., 7 Jul 2025). It proposes context-aware detection, multi-source data fusion, real-time lightweight methods, privacy-preserving techniques, and edge–cloud adaptive architectures. Its architectural components include collectors for metrics, logs, traces, billing or usage counters, and event-source metadata; modality-specific detectors; a fusion engine; and guardrail enforcers.
The review paper catalogs several detection systems and reported performance figures (Dorsett et al., 24 Aug 2025). Gringotts uses real-time monitoring of function performance metrics and Mahalanobis-distance anomaly detection, with reported performance of more than 95.75% accuracy, 1.86 s detection delay, and less than 1.1% overhead. DoWNet converts request sequences into temporal-spatial heatmaps and classifies them with a CNN, with 97.98% accuracy reported. FODWNN-DoWAD combines Pair Barracuda Swarm Optimization, Deep Wavelet Neural Networks, and HLCCO, with 99.05% accuracy reported. Provider-native unsupervised deep learning using CloudTrail, CloudWatch, and X-Ray is reported to detect all tested attack types with false alarm rate 0.003.
The same review identifies DoWTS as the principal simulator for safe, controlled emulation of DoW scenarios and labeled dataset generation (Dorsett et al., 24 Aug 2025). It supports both burst and slow-burn profiles, which is significant because real-world DoW is not restricted to volumetric spikes. The review emphasizes, however, that the field still suffers from scarcity of real-world DoW datasets, so many results depend on synthetic data and controlled experiments.
A recurring methodological point across these works is that spend itself must become a first-class detection signal. The anomaly-detection vision paper therefore proposes rate-of-spend and budget guardrails conceptually, while the unified taxonomy recommends tying conditions 5–6 to source attribution, traffic baselines, autoscaling events, billing anomalies, invocation counts, concurrency, and duration distributions (Nguyen et al., 7 Jul 2025). A plausible implication is that DoW detection is inherently multimodal: invocation counts alone are insufficient, but cost signals without lineage and trigger context are likewise incomplete.
6. Mitigation strategies, trade-offs, and unresolved problems
The 2021 materials describe several mitigation directions consistent with the paper’s early-stage objective of enabling future defenses: cost-aware throttling and circuit breakers, budgets and hard caps, per-tenant quotas, authenticated or verified triggers, upstream validation, anomaly detection on spend rate, backpressure on fan-out, idempotency and deduplication, retry-policy tuning, egress controls, and provider-level support such as spend alerts, hard per-day prepay caps, and auto-suspension when spend rate exceeds thresholds (Kelly et al., 2021). Because these are framed as consistent mitigation strategies rather than quoted protocol prescriptions, they should be interpreted as research-aligned guidance rather than a definitive standards list.
The review literature adds more operationally framed controls: API rate limiting and throttling, execution time limits, cost-management alerts, adaptive billing caps, native cloud-provider monitoring, per-tenant quotas, token buckets, and anomaly-informed throttling (Dorsett et al., 24 Aug 2025). These defenses differ from classical DoS measures because they are explicitly budget-aware. A notable trade-off repeatedly emphasized across the literature is that effective DoW mitigation may intentionally deny service to preserve sustainability. Hard caps and fail-closed spend policies can protect the wallet, but they can also interrupt legitimate bursts and violate service-level objectives.
The unified taxonomy aligns mitigation with observable conditions. It recommends cost-aware rate limiting, budget caps and policy-based throttling, anomaly detection on invocation counts and side-effect costs, burst buffers, pre-authorization and provenance checks, dynamic pricing guardrails, concurrency limits, reserved capacity, and verification of event sources such as trusted queues and buckets (Dorsett et al., 24 Aug 2025). Because DoW may arise from legitimate-looking events and trusted SaaS sources, controls must often be placed upstream of expensive processing rather than solely at HTTP ingress.
Several open problems recur across the surveyed papers. The anomaly-detection vision paper emphasizes robust cross-function correlation, context modeling across tenants and deployments, distinguishing legitimate bursts from attacks, concept drift, privacy–utility trade-offs, lightweight on-function detectors, data generation and labeling, and explainability tied to automated mitigation (Nguyen et al., 7 Jul 2025). The review highlights scarcity of real-world datasets, adversarial adaptation, and the need for adaptive billing models and provider-side safeguards (Dorsett et al., 24 Aug 2025). The unified taxonomy suggests future integration of socio-technical or behavioral dimensions into condition-based denial models (Dorsett et al., 24 Aug 2025).
A persistent misconception is that DoW is merely “DoS that costs money.” The literature does not support that simplification. DoW can overlap with availability failures, and DDoW or hybrid low-rate forms can satisfy additional conditions, but the defining target remains the victim’s budget and operational sustainability rather than immediate resource saturation (Dorsett et al., 24 Aug 2025). Another misconception is that ordinary autoscaling and uptime-preserving cloud design inherently mitigate denial attacks. In the serverless setting, those same mechanisms can become the attack multiplier (Kelly et al., 2021).
7. Broader uses of the term beyond serverless computing
Although serverless cost exhaustion is the dominant research meaning of Denial-of-Wallet, the term or its functional equivalent also appears in cryptocurrency and wallet-security contexts. The most direct cross-domain mapping is in MimbleWimble. “One Bad Apple Spoils the Bunch: Transaction DoS in MimbleWimble Blockchains” studies an attack in which an adversary on the Dandelion++ stem path aggregates a victim transaction with an attacker-crafted transaction and ensures the attacker’s transaction is mined instead. The synthesis explicitly characterizes this as a Denial-of-Wallet outcome because valid transactions fail to be included on-chain, and with 10% malicious nodes the adversary can prevent over 45% of all transactions from ending up in the blockchain (Tabatabaee et al., 2021). Here, the “wallet” being denied is the ability to complete a spend, not a cloud budget.
In hardware and secure-element wallet architecture, protective refusal to operate can also produce wallet unavailability. The crypto-terminal paper does not use the term DoW, but its threat model and countermeasure design translate failed attestation, cloning detection, firmware corruption, or unauthorized host interactions into deliberate refusal to sign or operate (Urien, 2023). This is a fail-safe form of denial: it preserves security by turning compromised states into intentional non-operation.
The broader cryptocurrency-wallet SoK places Denial-of-Wallet in an availability-focused taxonomy spanning social, device, application, network, protocol, smart-contract, custodial, and regulatory layers (Erinle et al., 2023). In that account, DoW includes permanent lockouts such as greedy smart contracts and destroyed libraries, temporary or selective censorship such as backend or mempool disruptions, and economic DoW in which fee dynamics price users out of spending. The paper’s scope is considerably wider than serverless DoW, but the shared conceptual core is denial of the wallet owner’s ability to access, authorize, or complete spends.
Browser-wallet analysis provides yet another neighboring usage. WalletRadar does not use the term DoW, and it primarily targets confidentiality and integrity risks, but its synthesis maps clickjacking, DOM-based XSS, redundant storage, and demonic vulnerability to wallet-specific availability failures such as inability to unlock, persistent UI obstruction, or broken operational flows (Xia et al., 2024). This broader family of wallet-denial phenomena should not be conflated with the serverless cloud threat, yet it demonstrates that the phrase “Denial-of-Wallet” has acquired multiple domain-specific meanings.
Taken together, the literature supports a two-level interpretation. In its strict and most developed form, DoW is the serverless cost-exhaustion attack class introduced in 2021 and elaborated by later taxonomies, reviews, and anomaly-detection work (Kelly et al., 2021). In a broader wallet-security sense, the term can denote denial of the wallet’s practical function, whether through attestation lockout, transaction non-inclusion, smart-contract freeze, or browser-wallet flow failure (Tabatabaee et al., 2021). The common denominator is not a single mechanism but the denial of economically or operationally meaningful wallet function.