DOSA: A Polysemous Acronym in Research

Updated 10 July 2026

DOSA is a polysemous acronym in research with domain-specific meanings spanning wireless security, DNN accelerator design, document annotation, optimization, continual learning, and cultural NLP.
In wireless security, DoSA quantifies the efficiency of synchronous key generation, while in accelerator design, DOSA employs gradient-based, one-loop optimization to dramatically improve energy-delay products.
DOSA also underpins innovations in automated document annotations, biomarker selection, and multi-label learning, highlighting the need for precise, domain-dependent disambiguation.

“DOSA” is a polysemous acronym in contemporary research literature rather than a single settled term. In recent arXiv usage, it denotes, among other things, Degree-of-Synchronous-Approaching in information-theoretic wireless security, Differentiable Model-Based One-Loop Search for DNN accelerators, Document Specific Automated Annotations for business documents, Dual-stage Optimizer for Systematic overestimation Adjustment in biomarker selection, Dual Output Spiking Architecture for continual multi-label learning, and Dataset of Social Artifacts for culturally grounded LLM evaluation. Capitalization is also unstable: some papers use DOSA, others DoSA, and several adjacent literatures use the related acronym DOS rather than DOSA (Sun et al., 2023).

1. Acronymic scope and nomenclature

Expansion	Research area	Representative paper
Degree-of-Synchronous-Approaching	wireless secrecy metrics	(Sun et al., 2023)
Differentiable Model-Based One-Loop Search	DNN accelerator co-design	(Hong et al., 12 Sep 2025)
Document Specific Automated Annotations	document intelligence	(Shukla et al., 2022)
Dual-stage Optimizer for Systematic overestimation Adjustment	biomarker selection, MO optimization	(Cattelani et al., 2023)
Dual Output Spiking Architecture	continual multi-label learning	(Mishra et al., 2024)
Dataset of Social Artifacts	cultural benchmark for LLMs	(Seth et al., 2024)

The multiplicity of meanings is not incidental. In some cases the acronym is formally introduced by the authors themselves, as with Degree-of-Synchronous-Approaching and Differentiable Model-Based One-Loop Search. In other cases, the literature explicitly notes nearby ambiguity: the deep-imbalance paper introduces DOS for “Deep Over-sampling,” while observing that broader literature may refer to “Deep Oversampling Approaches/Algorithms” as DOSA, although the paper itself does not (Ando et al., 2017). Similarly, the distributed storage paper uses DOS for “Distributed Object Store,” and the wireless MAC paper uses DOS and ADOS for distributed opportunistic scheduling rather than DOSA (Primmer, 2013, Garcia-Saavedra et al., 2014).

A recurring editorial issue is therefore acronym resolution by domain. In security and networking, DoSA may denote either a secrecy metric or a denial-of-sleep attack; in machine learning, it may denote an optimizer, a spiking architecture, or a dataset; in accelerator design, it denotes a co-design framework. This suggests that the term is best treated as a domain-dependent label rather than a singular concept.

2. DoSA as Degree-of-Synchronous-Approaching

In wireless information-theoretic security, DoSA stands for Degree-of-Synchronous-Approaching, introduced to measure how efficiently in time a system can approach Shannon’s one-time pad perfect secrecy (Sun et al., 2023). The formal definition is

$d = \frac{C_K}{C_M},$

where $C_K$ is the secret-key capacity and $C_M$ is the channel capacity. In this formulation, $d=1$ means secret keys can be generated as fast as data are transmitted, so one-time-pad-like operation is, in principle, synchronous; $d \ll 1$ means the system is key-rate limited.

DoSA was proposed as a complement to DoA (Degree-of-Approaching),

$D = \frac{E_K}{E_M},$

where $E_K$ is key entropy and $E_M$ is message entropy. The distinction is explicit: DoA measures closeness in security strength, whereas DoSA measures closeness in synchronous operation. A system can therefore exhibit high DoA but low DoSA if it eventually uses enough fresh key material yet generates that material too slowly for real-time use.

The paper develops this metric across two architectural paths toward Shannon’s one-time pad (Sun et al., 2023). In the explicit-encryption-based path, the central question is whether secret-key generation can keep pace with transmission. Channel-reciprocity-based key generation is described as having “an extremely low DoSA” because the key generation rate is much lower than the information transmission rate. By contrast, the proposed noise-based key generation (NBKG) architecture uses natural noise, full-duplex exchange, constellation rotation, and artificial noise to raise the secrecy key rate. Under high-SNR assumptions, its DoSA is approximated by

$d \approx 1 - \frac{2.83}{\log \mathrm{SNR}},$

so $d \to 1$ asymptotically as $C_K$ 0 increases.

In the implicit-encryption-based path, no explicit key is shared. Instead, physical-layer security and randomness shaping produce a virtual key from Eve’s uncertainty. In that setting,

$C_K$ 1

so DoSA numerically coincides with DoA, and the security condition is linked to Eve’s error floor $C_K$ 2 through

$C_K$ 3

The paper’s interpretation is direct: one-time-pad perfect secrecy corresponds to both $C_K$ 4 and $C_K$ 5, whether the key is explicit or virtual (Sun et al., 2023).

3. DOSA as Differentiable Model-Based One-Loop Search

In DNN accelerator design, DOSA denotes Differentiable Model-Based One-Loop Search for DNN Accelerators (Hong et al., 12 Sep 2025). The framework addresses joint hardware–mapping design-space exploration by making the accelerator performance model differentiable and then applying gradient-based optimization to hardware parameters and per-layer mappings in a single loop. Rather than using the conventional outer-loop hardware search plus inner-loop mapping search, it is explicitly “mapping-first, one-loop.”

The framework targets layers expressed over the canonical dimensions $C_K$ 6, parameterizes spatial and temporal tiling factors across memory levels, derives the minimal supporting hardware from those mappings, and optimizes the full-network energy-delay product

$C_K$ 7

Its analytical model gives closed-form expressions for traffic, compute latency, memory latency, and energy, and can be augmented with a learned correction model for real hardware. Reported results show 2.80× energy-delay-product improvement over random search and 12.59× over Bayesian optimization given a similar number of samples, together with 1.82× energy-delay-product improvement on real Gemmini RTL when optimizing buffer sizes and mappings with an analytical+learned model (Hong et al., 12 Sep 2025).

A subsequent accelerator paper positions DOSA as a state-of-the-art, data-driven performance predictor and HW/SW co-design framework used as a baseline for Polaris and Starlight (Sakhuja et al., 2024). In that comparison, DOSA is characterized as an offline differentiable proxy-model optimizer trained on RTL labels, exploring a smaller design space than Polaris because it does not explore spatial array dimensions and only explores three possible loop orders. Polaris is reported to match DOSA’s six-hour designs in under 35 minutes and to reduce energy-delay product by 2.7× over DOSA’s best designs in under 3.3 hours (Sakhuja et al., 2024). The juxtaposition is notable: one paper introduces DOSA as the optimization framework itself, while another treats it as the principal offline baseline for a multi-fidelity successor.

4. DoSA as Document Specific Automated Annotations

In document intelligence, DoSA stands for Document Specific Automated Annotations (Shukla et al., 2022). It is a human-in-the-loop system for generating initial annotations for business or form-like documents by bootstrapping from document-generic datasets and models. The motivating claim is that generic document models and generic labels such as “question/key,” “answer/value,” “header,” and “other” are insufficient for downstream business workflows that require document-specific fields such as invoice numbers, dates, or fax numbers.

The pipeline begins with OCR using Tesseract to obtain word-level text and bounding boxes, then applies LayoutLMv3 fine-tuned on FUNSD for generic key/value detection. The reported F1 on FUNSD for this generic entity classification is 0.9078 (Shukla et al., 2022). After generic key and value regions are detected, DoSA uses a heuristic key–value linking procedure rather than state-of-the-art form entity linking models, because those models produced “a lot of noisy pairs” in the reported tests.

The heuristic uses three conditions. For each value region $C_K$ 8, the system searches for a key region $C_K$ 9 such that: the position of $C_M$ 0 is less than the position of $C_M$ 1; $C_M$ 2 is not already linked to another value; and among keys satisfying the first two constraints, $C_M$ 3 is the closest to $C_M$ 4. The linked pair $C_M$ 5 is then converted into a document-specific annotation by assigning the text of the key as the label of the value region. A human reviewer corrects or normalizes these automatically generated labels, and the reviewed data are used to fine-tune a document-specific model. Subsequent iterations reuse that model to annotate more documents, again with human review, and the paper explicitly frames later-stage improvement in terms of active-learning ideas such as uncertainty-based sampling (Shukla et al., 2022).

The scope is stated to be limited to form-like documents, and the paper explicitly notes that empirical evaluation of annotation-time reduction and overall effectiveness is future work. It also states that an open-source ready-to-use implementation is available on GitHub (Shukla et al., 2022).

5. DOSA in optimization and continual learning

In omics-based biomarker selection, DOSA stands for Dual-stage Optimizer for Systematic overestimation Adjustment, and DOSA-MO is its multi-objective instantiation (Cattelani et al., 2023). The problem setting is the winner’s curse in model selection: when many feature subsets are evaluated, the best internal validation scores are systematically overestimated. DOSA-MO addresses this during optimization rather than after selection. For each objective $C_M$ 6, it models overestimation as

$C_M$ 7

learns a regression model

$C_M$ 8

and optimizes with adjusted fitness

$C_M$ 9

The regressors considered include dummy, ptree, RFReg, SVR, and rSVR, and evaluation is based on metrics such as MOPE, $d=1$ 0, and CHV. The reported finding is that DOSA-MO improves the performance of a state-of-the-art genetic algorithm on left-out or external sample sets for cancer subtype prediction and survival prediction in kidney and breast cancer transcriptomics datasets (Cattelani et al., 2023).

A different machine-learning usage appears in continual multi-label learning, where DOSA means Dual Output Spiking Architecture (Mishra et al., 2024). This architecture introduces positive and negative output heads for each label and defines the signed margin

$d=1$ 1

It is trained with an imbalance-aware focal maximum margin loss

$d=1$ 2

where $d=1$ 3 is a trainable per-class margin vector, and the paper also proposes an inverse-weighted F1 score

$d=1$ 4

to emphasize minority labels. Experiments on several benchmark multi-label datasets show improved robustness to data imbalance and better continual multi-label learning performance than CIFDM (Mishra et al., 2024).

These two uses share only the acronym. One is a dual-stage wrapper for multi-objective optimization under performance-estimation bias; the other is a spiking neural architecture with a dual-head output and imbalance-aware margin learning.

In NLP and cultural evaluation, DOSA stands for Dataset of Social Artifacts from Different Indian Geographical Subcultures (Seth et al., 2024). It is explicitly described as not about the food “dosa”, but about a community-generated benchmark for evaluating whether LLMs can recognize culturally salient objects and practices across Indian regional subcultures.

The dataset was built through participatory methods involving 260 participants from 19 different Indian geographic subcultures and contains 615 social artifacts (Seth et al., 2024). Artifact collection proceeded through surveys and a gamified “taboo”-style framework in which paired participants from the same state alternated between clue giving and guessing. The clue giver could provide at most five clues, could not use words synonymous with the artifact, and was instructed to emphasize information that most people from the same culture would know or agree with and that was most defining and distinctive of the artifact. This structure was intended to produce descriptions that “semantically align with the shared sensibilities” of the relevant communities.

The benchmark evaluates four popular LLMs—GPT-4, PaLM 2, Llama 2-13B Chat, and Falcon-7B—on an artifact-guessing task with accuracy@GUESS1, accuracy@GUESS2, and overall accuracy (Seth et al., 2024). The paper reports substantial variation across regional subcultures, stronger performance for closed-source models than for the open-source baselines, and a marked drop on the “expanded” artifact set collected after the game. It also reports that models often default to more popular or web-visible artifacts even when state context and clues indicate a less globally visible but locally salient answer. The broader implication drawn by the paper is that web-based training data produce geographically and socio-economically uneven cultural familiarity.

7. Peripheral usages, near-homographs, and recurrent confusion

Several adjacent literatures further complicate the interpretation of “DOSA.” In wireless sensor networks, DoSA commonly abbreviates Denial-of-Sleep Attack, an energy-exhaustion attack that prevents nodes from entering sleep mode. A representative countermeasure paper proposes a hybrid mobile-sink, firefly, and Hopfield-neural-network approach and evaluates it using metrics such as packet delivery ratio, average throughput, detection ratio, and network lifetime (Fotohi et al., 2020). In web systems, the XDoser paper does not formally use DOSA as its acronym, but it explicitly frames the tool as a denial-of-service-oriented assessment mechanism for synthetic server-load measurement, centered on HTTP POST flooding and metrics such as the failure-to-success ratio $d=1$ 5 (Haque et al., 2019).

Other confusions arise from papers that use DOS rather than DOSA. “Distributed Object Store Principles of Operation” uses DOS for distributed object store and discusses a distributed object store as software built on top of a distributed database, with a flat namespace, long time horizon, and capacity-on-demand requirements (Primmer, 2013). “Deep Over-sampling Framework for Classifying Imbalanced Data” introduces DOS for deep over-sampling in CNN feature space and explicitly notes that the paper itself uses DOS, not DOSA (Ando et al., 2017). “Adaptive Mechanism for Distributed Opportunistic Scheduling” develops DOS and ADOS for MAC-layer opportunistic scheduling rather than DOSA (Garcia-Saavedra et al., 2014).

A common misconception is therefore to treat DOSA as a uniquely identifiable method or benchmark. The literature instead shows a family of unrelated acronym expansions distributed across security, systems, machine learning, document intelligence, and cultural NLP. A plausible implication is that any technical use of “DOSA” requires immediate contextual disambiguation by field, capitalization, and citation.