Usman Rabiu ISAH
Laurent BOBELIN
Pascal BERTHOMÉ
INSA Centre Val de Loire Bourges, France Sule Lamido University Kafin Hausa, Nigeria [email protected]
INSA Centre Val de Loire Bourges, France [email protected]
INSA Centre Val de Loire Bourges, France [email protected]
Keywords
YARA rules are widely shared across threat intelligence communities to enable collective defence against malware. This practice implicitly assumes that removing metadata (e.g., author fields) sufficiently protects the identity of contributing organisations. To assess the validity of this assumption, we systematically evaluate how much can be inferred from YARA rule text alone. Specifically, using a corpus of 23,305 rules from three major public repositories, we train independent classifiers along four stylometric fingerprint dimensions: individual author, source repository, malware family, and temporal drift, using three complementary methods: lexical n-grams (Burrows’ Delta), syntactic AST features (Caliskan-Islam), and fine-tuned CodeBERT. Our results demonstrate that repository origin is almost perfectly recoverable (up to 99% accuracy), individual authors can be re-identified well above chance (76%), and malware family classification reaches 95%. Comparing the same repository attribution task across full-history and time-restricted subsets reveals a 9–18% accuracy gap, providing preliminary evidence of temporal drift in repository fingerprints. To further disentangle content from style, we conduct per-malware family author attribution experiments. Even when the malware family is the same for all samples considered, authors can still be re-identified for five of seven tested families (mean accuracy 74.6%). These findings constitute the first systematic demonstration that YARA rule sharing is a measurable OPSEC attack surface, and that metadata removal alone does not mitigate it.
YARA rules, authorship attribution, re-identification, stylometric fingerprinting, threat intelligence sharing, CodeBERT
20 26
Abstract
1
CC
S
ACM Reference Format: Usman Rabiu ISAH, Laurent BOBELIN, and Pascal BERTHOMÉ. 2018. Anonymous YARA Rules Are Not Anonymous. In Proceedings of Make sure to enter the correct conference title from your rights confirmation email (Conference acronym ’XX). ACM, New York, NY, USA, 14 pages. https://doi.org/ XXXXXXX.XXXXXXX
Introduction
Signature-based intrusion detection systems remain a cornerstone of operational cybersecurity. YARA is among the most widely used signature languages. It is a pattern-matching language for malware identification and classification [3]. YARA rules are routinely exchanged through public repositories, Information Sharing and Analysis Centres (ISACs), and inter-organisational partnerships to strengthen collective defence. This sharing is no longer just a best practice, it is now a regulatory expectation. The European Union’s NIS2 Directive (Directive 2022/2555) entered into force in October 2024. It requires cooperation, information sharing, and coordinated incident response. This mandate applies to operators of essential services in 18 critical sectors [11]. At the European level, sector-specific ISACs help exchange threat intelligence, including detection signatures, among their members. These include the European Energy ISAC (EE-ISAC), the Financial Institutes ISAC (FI-ISAC), the European Rail ISAC (ERISAC), and the Maritime ISAC [10]. National agencies also publish detection rules in threat advisories, and serve as a trusted intermediary in an ISAC: if one of the ISAC member identify a new threat, the trusted intermediary is in charge of anonymizing the detection content (IoC, detection rules) before sharing it to other partners, because those detection content is sensitive data. On the technical side, tools like OpenCTI, a platform co-developed by ANSSI, CERT-EU, and the Luatix non-profit, provides infrastructure for storing and disseminating YARA, Sigma, and STIX indicators across organisational boundaries [15]. In these sharing mechanisms, a common assumption prevails: that removing explicit metadata fields (e.g., the author tag, organisational names) from a rule before publication is sufficient to protect the contributor’s identity. The present work directly challenges this widespread assumption, setting the stage for our core analysis. We show that YARA rules carry persistent stylistic and structural fingerprints that survive metadata stripping. These fingerprints can be exploited by machine learning classifiers. An adversary can then recover the source repository of a rule with near-perfect accuracy.
AC M
to
ed
itt
CCS Concepts
bm
• Security and privacy → Pseudonymity, anonymity and untraceability; Privacy protections; Social aspects of security and privacy; Intrusion/anomaly detection and malware mitigation; Intrusion detection systems; Malware and its mitigation; • Computing methodologies → Supervised learning by classification; Machine learning.
Su
arXiv:2605.26791v1 [cs.CR] 26 May 2026
Anonymous YARA Rules Are Not Anonymous
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. Conference acronym ’XX, Woodstock, NY © 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 https://doi.org/XXXXXXX.XXXXXXX
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
20 26
Stage 1: Authoring and local conventions. A YARA rule begins life when an analyst encodes the distinguishing features of a malicious artefact as a signature. Even when targeting the same threat, two analysts produce substantively different rules: they select different strings, choose different variable naming schemes ($s1, $str_payload, $marker_01), decompose boolean logic differently, and prioritise different indicators. These choices reflect each analyst’s training and preferences, producing stable conventions across their corpus that aggregate into an institutional fingerprint when an organisation enforces shared templates or review checklists.
S
Stage 2: Tool-assisted generation and templating. Most productiongrade rules are not written from scratch. Generators such as yarGen [20] parse a malware sample, filter candidate strings against a goodware corpus, and emit a templated skeleton that the analyst then refines. The resulting tool signature is largely independent of the targeted malware: a yarGen-generated ransomware rule and a yarGengenerated RAT rule share the same skeleton, naming scheme, and condition structure, so any classifier that learns the skeleton attributes rules to the generator regardless of their target.
Su
bm
itt
ed
to
AC M
Contributions. This paper makes the following contributions: (1) We provide the first study of multi-dimensional stylometric fingerprinting of YARA rules, identifying candidates dimensions and determining if those dimension are attributable. Our empirical results shows YARA rules can carry multiple independent fingerprint dimensions: author, repository, and malware family. We compare different standard classification methods to identify those fingerprint dimensions: lexical n-grams, syntactic AST features, and fine-tuned CodeBERT. To determine if a dimension is part of the fingerprint, we use a confound-controlled author attribution experiment in which malware family and then both family and repository are held constant. (2) We also found preliminary evidence of temporal drift in repository fingerprints: on the one repository in our corpus with reliable per-rule timestamps (Neo23x0/signaturebase), attribution accuracy on the 2022–2025 subset reaches 99%, compared with 90% on the full history for our strongest method (CodeBERT), and drops to 81–83% for the handcrafted baselines. We frame this as a suggestive finding that motivates further study on corpora with richer temporal annotation, rather than as a definitive characterisation of convergence across the YARA ecosystem. (3) We provide a reproducible experimental pipeline and dataset methodology for future work in detection rule anonymity. To the best of our knowledge, this is the first systematic study of authorship and provenance attribution on YARA rules, and the first to demonstrate an externally verifiable institutional attribution chain in the detection-rule ecosystem. The remainder of this paper is organised as follows. Section 2 describes the YARA rule production pipeline and the fingerprints dimensions it introduces. Section 3 reviews related work in sourcecode authorship attribution and threat-intelligence sharing. Section 4 formalize the problem we study. Section 5 describes the dataset, features, and classification methods. Section 6 presents results along the four stylometric fingerprint dimensions we identified. Section 7 discusses implications, and Section 8 concludes.
repeatedly before it is shared. Figure 1 summarises this generic pipeline. The remainder of this section walks through each of its five stages and explains why each one can embed persistent structural or stylistic artifacts that survive metadata removal.
CC
In many cases, the individual author can also be identified at rates well above chance. Threat model. We consider a passive adversary who collects anonymised YARA rules from public repositories or ISAC feeds. The threat is particularly acute for organisations that contribute rules through ISACs or government-coordinated sharing programmes, where contributor anonymity and unlinkability of a producer’s rules are explicit design goals of the sharing framework (cf. NIST SP 800150’s principle of limited attribution [17]).
2
How YARA Rules Are Produced
To understand why shared YARA rules can leak the identity of their producer, we first describe how such rules come into existence. Detection rules are not static, hand-crafted artefacts. In practice they are the product of a multi-stage lifecycle that begins with the observation of malicious activity and ends with deployment and maintenance in operational environments [20]. Along the way, a combination of human analysts, automated generators, organisational policies, and format-conversion tools transform the rule
Stage 3: Normalisation and quality gates. Before acceptance into a repository, rules typically pass through linters and reformatters that enforce uniform indentation, consistent operator spacing, canonical hex casing, and prescribed keyword ordering. These gates replace individual variability with repository-wide uniformity: every rule that passes through the same gate emerges structurally similar to every other rule from that gate. Stage 4: Translation and backend adaptation. Rules are frequently rewritten for specific backends (VirusTotal Livehunt, endpoint agents, sandbox pipelines): regex expansion, PE-module substitution, condition-guard insertion, module renaming. Because organisations adopt consistent backends, these rewrites are repeated identically across many rules, imprinting a backend-specific pattern on top of stages 1–3. Stage 5: Distribution and sharing. The rule is published through a repository, ISAC feed, or bilateral exchange, typically after meta:block sanitisation. The artefacts introduced by stages 1–4 are preserved inside the rule body, clause ordering, condition nesting, naming conventions, boilerplate, because they are required for the rule to remain syntactically valid. Metadata removal thus leaves the stylometric surface intact. Implications. The cumulative effect of these five stages is that every shared YARA rule carries a layered fingerprint: individual authorial habits at stage 1, tool signatures at stage 2, repository conventions at stage 3, backend adaptations at stage 4, and residual formatting at stage 5. In practice, some of those dimension cannot be separated: human and tools working together at phase 1 and 2, in a co-authoring manner. On the other extremity of the pipeline, stage 3, 4 and 5 are hardly distinguishable: all those steps are subparts of the production and sharing rules policy of an organization, and thus, are not independent dimensions. We therefore hereafter
Anonymous YARA Rules Are Not Anonymous
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Attacker observes shared rules
Authoring & Local Conventions
Tooling & Templates
Normalization & Quality Gates
Translation & Backend Adaptation
Distribution & Sharing
Stylistic preferences (granularity, naming)
Tool fingerprints (boilerplate, ordering)
Standardized structure (consistent patterns)
Systematic rewrites (mappings, expansions)
Observable outputs (shared artifacts)
3 Related Work 3.1 Source Code Authorship Attribution
S
RF with LSTM, RNN, and CNN architectures, finding that deep learning consistently delivered superior performance [1]. Ullah et al. further advanced feature extraction by merging Program Dependence Graphs (PDGs) with deep learning to capture data and control-flow dependencies [21].
CC
Transformer era. Recently, the field has adopted pre-trained transformer models. Dipongkor et al. performed a comprehensive empirical investigation on models such as CodeBERT, finding these models not only highly effective but also considerably more resilient to adversarial attacks compared to traditional ML and DL methods [9]. Alvarez-Fidalgo et al. introduced CLAVE, a Transformer Encoder employing contrastive learning for authorship verification [4]. Choi et al. investigated zero-shot prompting with general-purpose LLMs (GPT-4) for attribution [7], though this approach faces validity concerns regarding data contamination, LLMs trained on opensource code may recall code-author pairs rather than learn genuine stylistic features. Our approach, using fine-tuned CodeBERT with controlled train/test splits, avoids this confound. This body of work traces a clear progression: from lexical ngrams to structural ASTs, and from traditional ML to transformers. However, prior research has predominantly focused on generalpurpose programming languages (C++, Java, Python) from academic contests or large-scale open-source projects. [16] To the best of our knowledge, the specialised domain of cybersecurity detection rules (YARA, Sigma, Snort, . . . ) has not been systematically investigated for attribution risk.
AC M
only two dimensions, that we name hereafter author and repository. Authors groups human authors and tool they used, and repository groups the end of the pipeline. In addition to those dimensions, the malware family the YARA rules detect is also a dimension of the fingerprint1 . The central question of this paper is whether these layers are strong enough, individually or in combination, to reidentify the producer of an anonymised rule. As any of those layers signature cand drift with time, we also studied if such a drift is identifiable, using methods and techiques coming from stylometric field. Section 3 reviews the stylometric and threat-intelligence literature related to our work.
20 26
Figure 1: Generic rule transformation and deployment pipeline.
ed
to
Source code authorship attribution (SCAA) rests on the principle of stylometry: that programmers exhibit a unique and distinguishable “fingerprint” in their coding style, analogous to a writer’s literary style [18]. Authorship attribution holds significant importance in cybersecurity for attributing malware and in software forensics for addressing plagiarism, intellectual property theft, and copyright disputes [1, 18].
Su
bm
itt
Feature engineering era. Early work focused on engineering features from code text, divided into lexical and layout categories. Frantzeskou et al. demonstrated that byte-level n-grams could build effective “author profiles,” proving the approach viable even with limited and comment-stripped code samples [14]. A significant advancement came from Caliskan-Islam et al., who introduced features derived from Abstract Syntax Trees (ASTs). By combining structural features with lexical and layout features, they used a Random Forest classifier to de-anonymise programmers at scale, firmly establishing the privacy and security implications of SCAA in public repositories [6]. Deep learning transition. As the field progressed, research transitioned from conventional machine learning (SVMs, Random Forests) to deep learning. Alsulami et al. were among the first to show that LSTM networks trained on AST-derived features could significantly surpass models with manually crafted features [2]. Comparative studies validated this pattern: Frankel and Ghosh assessed both logistic regression and deep learning comprising NLP-derived features alongside ASTs [13], while Alalawi et al. compared SVM and 1We determined empirically that malware alone does not carry enough information, while the family it belongs to is reattributable well above chance.
3.2
Threat Intelligence Sharing and Its Risks
The sharing of cyber threat intelligence (CTI), including detection signatures, is a well-established practice supported by formal standards such as STIX and TAXII [22]. At the institutional level, ISACs provide trusted communities where sector-specific organisations exchange threat data using protocols like the Traffic Light Protocol (TLP) [10]. In Europe, the NIS2 Directive has further reinforced this model by mandating cooperation and information sharing across critical infrastructure operators [11]. National agencies such as France’s ANSSI (through CERT-FR) and the US CISA actively publish detection rules as part of their threat advisories [5]. Despite this widespread sharing, the privacy implications of sharing detection rules have received comparatively little attention [8]. The Equation Group analysis by Kaspersky [19] demonstrated that persistent code-level stylistic characteristics can identify advanced threat actors, suggesting a symmetric risk for defenders sharing
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
This section formalises the threat model previewed in Section 1 and defines the attribution tasks we evaluate.
Research questions. Three research questions are addressed in this paper about those attribution tasks: RQ1 (Feasibility): Can each of the attribution tasks above be solved substantially above random baseline on anonymised YARA rules, and what are the best methods to do that? RQ2 (Temporal drift): Does the accuracy of 𝑓𝑝 on R timed differ from its accuracy on R, and if so, how? RQ3 (Style vs. content): When author attribution is performed within a single malware family (and optionally within a single repository), does the signal persist? If yes, the classifier is learning stylistic fingerprints rather than content. To answer those questions, we run experiment using the 3 main state-of-the-art classification methods (N-grams, AST, and transformers) and a confound-controlled experiment to disentangle the different dimension influencing attribution. The methodology we use is given in the next section.
AC M
Notations. Let R = {𝑟 1, 𝑟 2, . . . , 𝑟 𝑁 } be a corpus of anonymised YARA rules, where each 𝑟𝑖 has been stripped of its meta: block, block comments, and tags (see Section 5.2). Each rule carries three ground-truth labels drawn from the production pipeline of Section 2: its individual author 𝑎𝑖 ∈ A, its source repository 𝑝𝑖 ∈ P, and the malware it targets 𝑚𝑖 ∈ M. An adversary observes only the anonymised rule body 𝑟𝑖 and seeks to recover one or more of these labels. Note that the author label from YARA actually embed two dimensions identified earlier: human author and tools, as we don’t have access to the author tool setup. Some of the tools cite themselves as author in the corresponding field. Human authors are free to discard the tool’s name and replace it by their name, or to keep them as co-author of the tool. In this later case, that represents more than a third of our dataset, we considered the tool as a co-author of the rule. To study how attribution risk evolves over time, we additionally construct a time-restricted subset R timed ⊂ R containing only rules authored during 2022–2025 in Neo23x0/signature-base (the only repository in our corpus that exposes reliable per-rule authoring timestamps through its Git history; see §5.1). Temporal drift is then operationalised as the difference in repository-attribution accuracy between R and R timed , rather than as a per-rule label.
20 26
Problem Statement
S
4
(3) Repository attribution (full history): train the same classifier 𝑓𝑝 on the full corpus R. Comparing the time-restricted and full-history variants isolates the effect of temporal drift in repository conventions. (4) Malware family classification: learn 𝑓𝑚 : 𝑟 → M to recover the threat category the rule targets. Security properties at stake. Following NIST SP 800-150’s principle of limited attribution [17], rule sharing should preserve two properties: producer anonymity (a shared rule should not be attributable to its producer beyond chance) and unlinkability (multiple shared rules should not be reliably associable with the same producer). Our attribution tasks test both: single-rule attribution violates producer anonymity, and the per-repository and per-family results violate unlinkability.
CC
their own detection logic. Yet no prior work has systematically examined whether detection rules as authored artefacts, as opposed to the malware they detect, carry exploitable fingerprints [23].
Su
bm
itt
ed
to
Threat model and attacker capabilities. We consider a passive adversary operating in the collaborative YARA sharing ecosystem (public repositories, ISAC feeds, CERT advisories). The adversary does not inject, modify, or delete rules, and does not compromise the sharing infrastructure. Its objective is to de-anonymise the producer of shared rules and, by doing so, infer sensitive information about the originating organisation. The adversary has access to (i) a public sample of anonymised rules with ground-truth labels, which it uses to train classifiers, and (ii) one unlabelled anonymised rule whose producer it seeks to identify. It does not have access to the original unredacted rules, repository commit histories, internal IDS telemetry, or the tools and templates used to generate the rules. The adversary operates only on the shared rule artefact itself, that is, its post-anonymisation rule body (strings, condition, inline comments). We assume no side channels (network timing, platform identifiers, infrastructure logs) and no auxiliary metadata.
5 Methodology 5.1 Dataset Construction We collected YARA rules from three major public GitHub repositories: Neo23x0/signature-base (maintained primarily by Florian Roth), reversinglabs-yara-rules (ReversingLabs), and YARAHQ/yaraforge. After parsing, deduplication, and metadata stripping, the unified corpus contains 23,305 rules spanning 67 authors and 3 repositories. Each rule carries author, repository, and malware-family annotations. All author, full-history repository, and malware-family experiments train on this unified corpus (yara4_with_repo.csv). Task-specific class-size filters are applied independently to this single source (§5.2).
Attribution tasks. We define four attribution tasks over the label set and subsets above, each corresponding to one of the fingerprint dimensions (identified as candidates in section 2, with the exception stated earlier that author and tool signatures cannot be distinguished in our dataset) introduced by the production pipeline:
Time-restricted subset. For the timed repository-attribution task, we additionally use yara2.csv, a subset comprising the 14,366 rules authored during 2022–2025 in Neo23x0/signature-base, the only repository in our corpus that exposes reliable per-rule authoring timestamps through its Git history. Comparing full-history and time-restricted repository attribution isolates the effect of temporal drift in repository conventions (§7.2).
(1) Author attribution: learn 𝑓𝑎 : 𝑟 → A to recover the individual analyst who wrote the rule, trained on R timed . (2) Repository attribution (time-restricted): learn 𝑓𝑝 : 𝑟 → P on R timed .
Malware-family labels. Family labels were assigned via a twostep pipeline: the rule_name and meta: description fields were extracted from each raw rule and submitted to a large language model (Claude, Anthropic) with a structured prompt requesting
Anonymous YARA Rules Are Not Anonymous
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Each rule undergoes a two-stage preprocessing step designed to eliminate explicit identifiers while preserving the stylometric surface. Stripping. We remove the meta: block in its entirety, as it contains the strongest explicit identifiers (author name, description, date, hash references). Block comments (/* . . . */) are also removed because they frequently contain copyright headers or analyst signatures. Rule-level tags (e.g., rule Name : tag) are stripped for the same reason.
Effect on stratified splits. Where multi-attribution rows exist, our 80/20 stratified split (Section 5.2) may place identical bodies under different labels into both training and validation sets. This affects the author-attribution task in Table 1 most directly; repository, malware-family, and timed repository tasks are largely unaffected, as their labels are not author-derived. The per-family confoundcontrolled experiments of Section 6.2 train within fixed (family, repository) subsets and substantially constrain the duplication pattern.
AC M
Retention. We deliberately retain the strings: section, the condition: block, and any inline comments. These three elements constitute the primary stylometric surface of a YARA rule: the choice of variable names, hexadecimal encoding preferences, string literal formatting, and the structure of boolean conditions are all degrees of freedom that analysts exercise differently. As established in Section 2, these are precisely the layers at which both individual authorial habits and tool-induced fingerprints are introduced and persist through the sharing pipeline.
20 26
Preprocessing and Anonymisation
Multi-attribution rules. Some rules in the upstream repositories appear with author strings that span multiple analysts. Our pipeline treats each distinct comma-separated author token as a separate label, which means a rule whose meta declares two authors generates two CSV rows with identical rule bodies and distinct labels. In our master corpus, 7,442 of 19,224 rows share a rule_name with at least one other row; 2,182 distinct rule names appear with byte-identical bodies under multiple author labels. We retain these rows because they reflect the true distribution of analyst attributions in published YARA corpora, which is the population a downstream attacker would observe; removing them would also remove a real signal about co-authorship patterns in the threat-intelligence community.
S
5.2
the upstream corpus and our pipeline interact in ways that warrant explicit discussion.
CC
classification into a fixed taxonomy of 22 families (e.g., APT, Backdoor, Ransomware, Webshell). Rules that could not be confidently categorised were assigned the label Other and excluded from the malware-family task, yielding 16 retained families. Note that the rule_name and meta: description fields are not included in the rule text presented to classifiers; they are used solely for label assignment and are stripped during preprocessing (§5.2).
Su
bm
itt
ed
to
Class-size filtering. For author attribution, we retain only authors with between 50 and 500 rules. The lower bound of 50 reflects the minimum number of samples needed to obtain a meaningful 80/20 stratified split (40 training rules / 10 validation rules per author); below this threshold, validation estimates become dominated by sampling noise and the model has insufficient signal to learn a stable stylistic profile. The upper bound of 500 addresses a pronounced long-tail imbalance in our corpus: a small number of prolific maintainers (notably Florian Roth, who authored 56% of the unfiltered dataset) would otherwise dominate the class distribution, allowing the classifier to achieve high apparent accuracy by learning a single dominant style rather than discriminating among authors. Capping at 500 rules per author yields a more balanced training set in which macro-averaged F1 becomes a meaningful indicator of per-author attribution risk rather than a reflection of the majority class. For repository and malware-family tasks, we retain all classes with at least 50 rules under the same minimum-signal rationale; no upper cap is applied, as these tasks have substantially fewer classes (3 and 16 respectively) and the class-imbalance effect is less severe. An 80/20 stratified train–validation split is applied uniformly across all tasks.
5.3
Corpus-level artefacts and leakage controls
YARA repositories are community resources in which rules are frequently re-used, refined, or re-attributed across analysts. Our extraction pipeline emits one CSV row per (rule, author) pair, where authorship is taken from each rule’s meta block. Two properties of
Identifier-leakage ablation. A second leakage channel, distinct from row-level duplication, is whether attribution accuracy is driven by analyst signatures, tool names, or URLs embedded in inline // comments. We address this directly in Section 7.4 by retraining both the author and timed-repository tasks on a corpus variant in which all inline comments have been stripped. Accuracy changes by at most 2 pp and moves in the direction opposite to the identifier-leak hypothesis (stripping comments slightly improves author attribution rather than degrading it), supporting the claim that the stylometric signal resides in the rule body and condition logic rather than in human-readable identifiers.
5.4
Classification Methods
We apply three complementary classifiers. Lexical n-grams. Following Burrows et al. [18], we extract character n-grams (𝑛 = 1, . . . , 6) from raw rule text and classify with a nearest-centroid classifier under Burrows’ Delta distance. This captures surface-level lexical habits. AST features. Following Caliskan-Islam et al. [6], we parse each rule into an abstract syntax tree and extract node-type frequencies, tree depth, branching factors, and node-type bigrams; classification uses a Random Forest with 300 estimators. This captures syntactic habits invariant to surface formatting. Fine-tuned CodeBERT.. We fine-tune microsoft/codebert-base [12] with a linear classification head, training one independent singletask model per attribution task. Author, full-history repository, and malware-family tasks train on the unified corpus; the timed repository task trains on the Neo23x0 2022–2025 subset. Rule text is tokenised and truncated to 512 tokens; training uses AdamW
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
Per-Family Confound Control
A key concern in author attribution of detection rules is content confounding: authors who specialise in specific malware families may be identifiable by the content they detect rather than their writing style. To disentangle these effects, we design a per-family experiment with two tiers: Tier 1: Per-Family. For each malware family with 3 or more authors, each having at least 20 rules, we train an independent CodeBERT classifier. It predicts the author within that family only. If attribution succeeds, the signal must come from style rather than content, since all rules in the training set target the same malware. Tier 2: Per-Family × Per-Repository. We further partition by repository, training author classifiers within each (family, repository) pair. This eliminates organisational coding standards as an alternative explanation, isolating individual stylistic fingerprints.
AC M
5.6
5.7
20 26
Stylometric Fingerprint Dimensions
Reporting convention. In all tables we report accuracy, M-F1, and W-F1 on the held-out 20% validation split, stratified by the target label. The gap between M-F1 and W-F1 provides diagnostic insight into whether a method relies on majority-class performance or generalises uniformly across classes.
S
5.5
We evaluate each method to perform the attribution tasks defined in section 4, with our dataset, i.e.: (1) Author Attribution (17 classes): Predict the individual author of a rule from the unified corpus, after applying the 50–500 rules-per-author class-size filter (§5.2). (2) Repository Attribution – Timed (3 classes): Predict the source repository from the same time-restricted subset. (3) Repository Attribution – Full (3 classes): The same repository prediction task on the full time-range dataset. (4) Malware Family Classification (16 classes): Predict the malware family a rule targets. Comparing Repository Attribution (Full) against Repository Attribution (Timed) isolates the effect of temporal drift in repository conventions: if conventions have become more standardised over the 2022–2025 window, we expect the timed subset to yield higher accuracy.
baseline is 1/𝐾; for 3-way repository attribution it is 33.3%; for 16-way malware classification it is 6.25%. We use this baseline as the reference point for the attribution-verdict thresholds defined in Section 5.7.
CC
(lr 2 × 10−5 , batch 16) with early stopping (patience 3) on validation accuracy. Experiments ran on PyTorch 1.10.2 (CUDA 11.1) and HuggingFace Transformers 4.30.0; the full environment is in the released artefact bundle. All experiments were run with PyTorch 1.10.2 (CUDA 11.1) and HuggingFace Transformers 4.30.0; the complete environment specification is included in the released artifact bundle.
Evaluation Metrics
Attribution verdicts. We classify each per-family experiment by the ratio of its observed accuracy to the random baseline 1/𝐾, where 𝐾 is the number of candidate authors in that experiment. Three verdict bands are defined:
Accuracy. The fraction of validation rules whose predicted label matches the ground truth. Accuracy is the simplest summary but is dominated by the majority class under imbalance.
Attributable (acc > 2 × 1/𝐾): the classifier substantially exceeds random guessing, indicating that stylistic fingerprints are strong enough to pose a concrete re-identification risk even under content or organisational control. Partially resilient (1.5 × 1/𝐾 < acc ≤ 2 × 1/𝐾): the classifier outperforms random guessing but not decisively; fingerprints exist but are attenuated by template use or style overlap. Resilient (acc ≤ 1.5 × 1/𝐾): the classifier is at or near random performance, indicating that the combination of controlled content, organisational standardisation, and limited stylistic variation suppresses identifiable fingerprints.
ed
to
Given the severe class imbalance across our datasets (see Section 5.1), reporting a single accuracy figure would obscure per-class performance. We therefore report three metrics throughout the paper.
bm
itt
Macro-averaged F1 (M-F1). F1 is computed independently for each class and then averaged with equal weight, regardless of class size: 1 ∑︁ M-F1 = F1𝑐 . (1) |𝐶 | 𝑐 ∈𝐶
Su
Because every class contributes equally, M-F1 is sensitive to performance on minority classes. From an OPSEC standpoint this is the more diagnostic metric: a single attributable author represents a concrete re-identification risk regardless of how few rules they have authored. Weighted F1 (W-F1). Per-class F1 scores are averaged weighted by the number of samples per class: ∑︁ |𝑆𝑐 | W-F1 = F1𝑐 . (2) |𝑆 total | 𝑐 ∈𝐶
The choice of 2× as the “clearly attributable” threshold reflects the OPSEC interpretation: if an adversary’s posterior belief about the author of a rule is at least twice the uniform prior, the rule has leaked meaningful information about its producer. The 1.5× intermediate band captures cases where the leak is statistically detectable but operationally marginal.
W-F1 reflects the expected outcome for a randomly sampled rule from the corpus and is dominated by the most prolific classes.
6 Results 6.1 Multi-Method Comparison
Random baseline. For a classification task with 𝐾 equally likely classes, the random-guessing baseline accuracy is 1/𝐾. For author attribution within a malware family containing 𝐾 authors, the
Table 1 presents the complete results for the three methods we used applied to the four fingerprint dimensions identified. All metrics are reported on held-out validation sets (20% stratified split).
Anonymous YARA Rules Are Not Anonymous
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Table 1: Multi-method, multi-axis attribution results. Accuracy (%), macro-averaged F1 (M-F1), and weighted F1 (W-F1) are reported. Best result per task in bold. N-grams
AST
CodeBERT
Task
Classes
Acc
M-F1
W-F1
Acc
M-F1
W-F1
Acc
M-F1
W-F1
Author Repo (Timed) Repo (Full) Malware
17 3 3 16
58 99 81 93
0.60 0.99 0.85 0.79
0.57 0.99 0.79 0.92
52 98 83 89
0.54 0.98 0.87 0.65
0.51 0.98 0.83 0.88
76 99 90 95
0.76 0.99 0.93 0.93
0.75 0.99 0.91 0.95
Rules
Acc (%)
3 5 4 9 3 19 4
358 310 339 514 285 2,144 632
91.7 88.7 88.2 77.7 70.2 61.3 44.1
Attributable Attributable Attributable Attributable Attributable Partial Resilient
74.6
—
CC
Ransomware Backdoor Exploit APT RAT Hacktool Webshell
Authors
S
Family
20 26
Table 2: Per-family author attribution with CodeBERT. Ran1 . dom baseline is authors
AC M
Mean (7 families)
to
Figure 2: Temporal convergence effect: repository attribution accuracy on the 2022–2025 subset vs. full history.
Su
bm
itt
ed
Second, repository attribution on the timed subset approaches ceiling performance (99% for both n-grams and CodeBERT), while the same task on the full dataset drops to 81–90%. This 9–18 percentage point gap is consistent with repository-specific conventions converging over the 2022–2025 period, which would make rules from this window more distinctively fingerprinted. Figure 2 isolates this temporal effect. We examine alternative explanations for the gap in Section 7.2, such as virus drift. Two caveats constrain the interpretation of this result. First, only Neo23x0/signature-base exposes reliable per-rule authoring timestamps through its Git history; ReversingLabs and YaraHQ/yaraforge provide rules without dependable temporal metadata. The “timed” subset is therefore effectively a Neo23x0-weighted slice of the corpus, and the gap should be read as a within-repository observation rather than a cross-ecosystem trend. Second, the 2022– 2025 window spans only four calendar years. We therefore treat this result as preliminary evidence that temporal drift in repository fingerprints is measurable, not as a definitive characterisation of how repository conventions evolve. Third, n-grams outperform AST features on content-heavy tasks (malware: 93% vs. 89%). This is expected: malware-specific string patterns (e.g., hexadecimal signatures) are captured by lexical n-grams but not by structural AST features. Conversely, AST features slightly outperform n-grams on the full repository task
Verdict
(83% vs. 81%), suggesting that structural conventions (e.g., condition complexity) are more stable across time than surface-level formatting.
6.2
Per-Family Author Attribution (Confound Control)
To estimate if author attribution good results may be due to other fingerprints dimensions, we performed authors attribution by splitting datasets by malware families, using only rules targeting a specific family for both training and testing (tier 1). Table 2 shows these author attribution results within each malware family. Five of seven tested families are classified as attributable, with Ransomware achieving 91.7% accuracy across 3 authors (random baseline: 33%). Only Webshell exhibits resilience to attribution (44.1% with 4 authors, baseline 25%), likely because webshell rules follow highly standardized structural templates that suppress individual variation. Hacktool, the largest and most diverse family (19 authors, 2,144 rules), falls in the partially resilient category at 61.3% (baseline 5.3%), still far above chance. Figure 3 visualizes these results against their random baselines. Table 3 shows Tier 2 results, i.e. results for each family and repository couple. Even after controlling for both malware family and repository, four of eight experiments remain attributable. Notably, Webshell × Neo23x0 is attributable at 77.0% while Webshell × YaraHQ collapses to chance (33.3%), suggesting that YaraHQ’s webshell rules follow such rigid templates that individual authorship signals are completely suppressed. This contrast provides direct evidence that organizational style conventions can either amplify or suppress individual fingerprints. Figure 4 presents these results visually.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
Table 4: Hacktool per-family accuracy as the per-author rule cap 𝑁 varies. 𝐾 is the number of qualifying authors at each cap; random baseline is 1/𝐾. 𝑁 =20
𝑁 =80
𝑁 =100
𝑁 =200
𝑁 =400
T1: Hacktool (𝐾 authors) T2: Hacktool×Neo23x0 (𝐾 authors) T2: Hacktool×YaraHQ (𝐾 authors)
50.0% (19) 68.8% (8) 48.1% (13)
57.5% (10) 53.1% (4) 53.9% (8)
60.6% (8) 53.8% (4) 62.5% (6)
63.2% (7) —
65.0% (6) —
60.0% (5)
49.7% (4)
20 26
Setting
Figure 3: Tier 1: Per-family author attribution accuracy.
Rules
Acc (%)
Verdict
Exploit × Neo23x0 APT × YaraHQ APT × Neo23x0 Webshell × Neo23x0 Backdoor × YaraHQ Hacktool × Neo23x0 Hacktool × YaraHQ Webshell × YaraHQ
3 3 5 3 3 8 13 3
254 303 324 431 121 937 1,497 600
94.1 80.3 78.5 77.0 68.0 67.0 54.7 33.3
Attrib. Attrib. Attrib. Attrib. Partial Partial Partial Resilient
S
Auth.
Accuracy trends. In the unrestricted Tier 1 setting, raw accuracy rises monotonically from 50.0% at 𝑁 =20 (19 authors, baseline 5.3%) to 65.0% at 𝑁 =400 (6 authors, baseline 16.7%). Because the baseline also rises as authors are filtered out, raw accuracy alone is misleading; the accuracy-to-baseline ratio is the meaningful quantity. That ratio remains consistently in the attributable band: 9.4× at 𝑁 =20, falling to 3.9× at 𝑁 =400. The downward trend in this ratio reflects the narrowing of the author set to the most prolific, and therefore most homogeneously “professional”, contributors, whose styles are harder to separate. The Tier 2 settings show less regular behaviour because the coupled filtering is more aggressive: for Hacktool×Neo23x0, only 4 authors retain at least 80 rules, which mechanically caps achievable accuracy at the level of the closest pair. This confirms that the fingerprinting signal observed in our headline results is robust to sample-size choices and is not an artefact of any single cap.
Su
bm
itt
ed
to
AC M
Family × Repo
how fingerprint signal concentrates as the per-author corpus grows and as the author set narrows to the most prolific contributors.
CC
Table 3: Tier 2: Per-family × per-repository author attribution. Controls for both content and organizational conventions.
Figure 4: Tier 2: Per-family × per-repository author attribution.
The fixed per-author cap used in Section 6.2 leaves one question unanswered: how does the amount of training data available per author shape what the classifier actually learns? To probe this, we repeat the Hacktool per-family experiment while sweeping a fixed per-author rule cap 𝑁 ∈ {20, 80, 100, 200, 400} across both tiers. Varying 𝑁 has two coupled effects. First, it changes the number of training samples available per author. Second, it implicitly changes the set of authors who qualify for inclusion, since authors with fewer than 𝑁 rules are excluded at each cap. This design exposes
An unexpected institutional signal. More striking than the accuracy trend is a qualitative pattern that emerges from the confusion matrices themselves. Across every Tier 2 Hacktool×Neo23x0 run, the classifier systematically confuses yarGen-generated rules with those written by Tobias Michalski (an employee of Nextron Systems, Nextron Systems being the company developping YarGen): 88% of yarGen-labelled rules are predicted as Michalski at 𝑁 =80 (Figure 5), 85% at 𝑁 =100, and a similar directional pattern (44% yarGen→Michalski) persists in Tier 1 at 𝑁 =80. The confusion is directional, yarGen is predicted as Michalski far more often than the reverse, and is stable across caps. It is not the kind of symmetric error one would expect from two genuinely similar authors; it has the profile of a many-to-one mapping in which the “yarGen” class is being absorbed into the “Michalski” class. Implication: institutional is easier than individual. This finding extends the attribution story in two ways. First, it demonstrates that the stylometric signal does not always decompose cleanly into “human author” versus “tool fingerprint.” When an analyst and a tool share a production environment, their fingerprints converge, and any classifier trained to separate them will instead recover the signature of the couple formed by them.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
CC
S
20 26
Anonymous YARA Rules Are Not Anonymous
AC M
Figure 5: Tier 2 Hacktool×Neo23x0 confusion matrix at 𝑁 =80 (row-normalised). Table 5: Co-attribution overlap among Nextron-affiliated authors in Hacktool × Neo23x0. Co-attributed rules
Institutional Co-authorship in the Nextron Pipeline
itt
6.3
174 172 172
ed
Michalski ↔ yarGen Florian Roth ↔ Michalski Florian Roth ↔ yarGen
to
Author pair
Nextron’s tooling and its named analysts is robust enough to be recovered by either method independently.
Generalisability. The Nextron triad is the most prominent example in our corpus but not unique: across the unified corpus, 3,514 distinct rule names appear under multiple author labels, of which 2,182 are byte-identical across attributions (Section 5.3). The co-authorship signal we describe is a structural property of how analyst-authored and tool-generated rules co-exist in publicly shared corpora.
7 Discussion 7.1 What Drives Attribution
Su
bm
The varying-𝑁 confusion described above admits a more concrete explanation than stylometric similarity: the upstream Neo23x0/signature- Our results establish a clear hierarchy of attribution signal strength base repository declares the institutional coupling in its meta blocks. across fingerprint dimensions in our dataset: repository > malware Of the 174 rules attributed to Tobias Michalski in Hacktool×Neo23x0, content > individual author > temporal period. Repository origin is all 174 share an identical rule_name with rules attributed to yarGen, the strongest signal (up to 99%), consistent with organisational codand 174 with rules attributed to Florian Roth. Inspection shows ing standards, naming conventions, boilerplate structure, condition the bodies are byte-identical (Table 5). The same content is pubtemplates, acting as a collective fingerprint. Individual author attrilished under multiple author labels, most commonly because rules bution, while lower in absolute terms (76%), is remarkable given originally generated by yarGen are subsequently curated and rethat it succeeds even within a single malware family (Tier 1 mean: attributed by analysts at the same organisation. Our extraction 74.6%) and within a single family-repository pair (Tier 2: four atpipeline (Section 5.3) preserves each declared attribution as a sepatributable pairs). The consistency across methods reinforces this rate row. hierarchy: even the weakest method (AST) achieves 83% on reposiThe classifier’s 88% Michalski→yarGen confusion is therefore tory attribution and 89% on malware classification, confirming that not the discovery of a hidden institutional voice; it is the recovthese signals are robust and not an artefact of any single feature ery of a multi-attribution pattern visible to anyone reading the representation. upstream meta blocks. The two views, model confusion and diA key insight from our per-author analysis is that attribution rect co-attribution counts, agree, which strengthens rather than vulnerability is not a simple function of data volume. The varyingweakens the institutional-attribution finding: the coupling between 𝑁 sweep in Section 6.2 shows that narrowing the author set to
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
Author (17 classes) Repository, timed (3 classes)
7.3
Baseline
–inline
Δ
76.4 99.2
78.4 99.3
+2.0 +0.1
The Content Confound and Its Resolution
The comparison between aggregate-level malware classification (95% for CodeBERT) and per-family author attribution (mean 74.6%) reveals the extent of content confounding in naive attribution setups. The 20-percentage-point drop when content is controlled confirms that a substantial portion, but not all, of the attribution signal in the unconstrained setting is driven by content rather than style. A classifier trained on the full dataset without family controls would conflate “Author X writes ransomware” with genuine stylistic attribution. Our per-family design eliminates this confound: within a single family, all rules target the same malware, so any above-chance accuracy must derive from stylistic differences. The Webshell × YaraHQ result (33.3% = random chance) illustrates the limiting case: when organisational templates fully standardise rule structure, individual fingerprints vanish. Conversely, Ransomware (91.7%) and Exploit × Neo23x0 (94.1%) show that where templates are less rigid, authorship signals are strong and persistent.
S
Temporal Drift: A Premise, Not a Conclusion
Task
Su
bm
itt
ed
to
AC M
The nine- to twenty-percentage-point accuracy gap between the timed and full-history repository tasks is consistent across all three methods, and the direction of the gap is the same in every case: the more recent the rules, the more attributable they become. Taken at face value, this is the signature one would expect if repository conventions were standardising over time, with newer rules passing through more uniform linting, templating, and review pipelines than older ones. However, our data cannot establish this conclusion on its own. The temporal signal rests on a single repository, Neo23x0/signaturebase, the only source in our corpus with reliable per-rule timestamps, and on a four-year window that is short relative to the decade-long history of public YARA sharing. The observed effect could equally reflect a Neo23x0-specific editorial shift, a change in its contributor base, or a change in what kinds of threats the community has prioritised during 2022–2025. We cannot, from our current experiments, distinguish between these hypotheses. We therefore present the temporal result as a premise for future work rather than as a finding. Two extensions are needed to sharpen it. First, the analysis should be replicated on repositories with richer temporal metadata, ideally spanning at least a decade, and with contributor sets that are stable across the window. Second, the attribution task itself should be decoupled from the filtering effect: controlling for the set of authors who are active in both sub-periods would isolate true convergence from a change in contributor composition. The method-specific variation in the gap (18 pp for n-grams, 15 pp for AST, 9 pp for CodeBERT) is nonetheless suggestive. If the effect were purely compositional it would be roughly uniform across methods; the fact that it is largest for surface-level features and smallest for deep contextual embeddings is consistent with the hypothesis that cosmetic standardisation has advanced faster than deeper structural convergence. This pattern is worth reporting precisely because it motivates the longer-horizon study needed to test it. For the purposes of the present paper, the temporal result should be understood as an invitation to that study, not as a completed one.
CC
7.2
Table 6: Effect of removing inline // comments on attribution accuracy. Accuracy is the best-epoch validation value. Baseline values correspond to Table 1.
20 26
only the most prolific contributors reduces the accuracy-to-baseline ratio from 9.4× at 𝑁 =20 to 3.9× at 𝑁 =400: when only high-volume contributors remain, their styles are more homogeneous and therefore harder to distinguish. This is the opposite of the intuition that “more data always helps the attacker” and has direct implications for contributor OPSEC, a prolific author who conforms to community conventions may paradoxically be harder to re-identify than a low-volume author with an idiosyncratic style. The varying-𝑁 analysis in Section 6.2 further refines this hierarchy: when tool-induced and individual fingerprints originate from the same organisational pipeline, they are not separable by a classifier, and the resulting signal is best understood as an institutional fingerprint rather than as either an individual or a tool one.
7.4
Ablation: Inline Comments
Because §5.2 deliberately retains inline // comments as part of the stylometric surface, a natural concern is whether attribution is driven by rule-body style or by analyst signatures, tool names, or URLs that may appear inside comments. We therefore rerun the two primary tasks on a variant of the corpus in which every inline comment has been removed, keeping the rest of the preprocessing pipeline identical. Table 6 reports the comparison. Accuracy changes by at most 2 percentage points in either direction, and does so in the direction opposite to what an identifier-leak hypothesis would predict: stripping comments marginally improves author attribution rather than degrading it. We conclude that the stylometric signal resides in the retained strings and condition blocks, not in analyst signatures or identifiers embedded in inline comments. Removing inline comments altogether would therefore be a defensible preprocessing choice; we retain them in the released corpus to preserve maximum fidelity to the shared-rule artefact as observed in practice.
7.5
Macro vs. Weighted F1: Reading the Results
Given the severe class imbalance across our datasets, we report both macro-averaged and weighted F1 scores. Macro F1 gives equal weight to every class regardless of size, making it critical for judging performance on minority classes, a single attributable author constitutes a real OPSEC risk regardless of how many rules they have written. Weighted F1 reflects the expected outcome for a randomly sampled rule.
Anonymous YARA Rules Are Not Anonymous
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
7.6
Implications for Practitioners
to
Repository-level fingerprinting is near-perfect. An adversary who observes a shared rule can, with high confidence, determine which major repository it originates from. For organisations that contribute exclusively to a single repository, this effectively de-anonymises the contributor.
itt
ed
Temporal analysis amplifies risk. Recent rules from standardised repositories are more attributable than historical ones. The trend toward rule quality standardisation, while beneficial for detection efficacy, inadvertently increases OPSEC exposure.
Su
bm
Countermeasures must be multi-layered. Both surface-level (ngram) and structural (AST) methods achieve above-chance attribution, so no single defence suffices: auto-formatting defeats ngrams but leaves structural signals intact, while restructuring conditions defeats AST features but preserves lexical habits. Effective approaches must address both layers, candidates include style transfer to a canonical form, federated detection (sharing models rather than rules), and differential privacy applied to rule structure. The Webshell × YaraHQ result (33.3%) shows that strict organisational templates can provide effective protection, at the cost of suppressing individual expertise.
7.7
Limitations
Several limitations should be noted. First, our dataset covers only three public repositories; private organisational rule sets may exhibit different characteristics. Second, the per-family experiments are limited to the seven families with sufficient multi-author representation; fifteen of 22
20 26
Conclusion
S
8
YARA rules are shared under the working assumption that stripping metadata is enough to protect the identity of the organisation that wrote them. To the best of our knowledge, this paper is the first to test that assumption empirically on YARA rules, at scale, and across four complementary fingerprint dimensions. While prior source-code authorship attribution work has focused on generalpurpose programming languages drawn from academic contests or large open-source projects, no prior study has systematically examined whether cybersecurity detection rules, YARA, Sigma, or Snort, carry exploitable fingerprints as authored artefacts distinct from the malware they detect. Our answer is unambiguous: the metadata-stripping assumption is wrong. Repository origin is almost perfectly recoverable (99% accuracy), individual authors can be re-identified well above chance (76%), and malware-family specialisation is nearly deterministic (95%). None of these signals depend on explicit identifiers; they live in the strings and condition blocks that any useful rule must keep. More importantly, the signal does not disappear under the controls that a reasonable defender would expect to work. Holding malware content constant still leaves five of seven families attributable; holding both content and repository constant still leaves four of eight configurations attributable. The attribution surface is multi-layered, and no single sanitisation step cuts through it. The most consequential result is not a number but a chain. By sweeping the per-author rule cap, we showed that the classifier systematically merges yarGen-generated rules into Tobias Michalski’s class at the 85–88% level, and that this “error” corresponds to a real, publicly verifiable institutional coupling: Michalski and Florian Roth are both affiliated with Nextron Systems, the company that also maintains yarGen and the Neo23x0 repository. What looks like a confusion between a tool and an analyst is in fact the classifier recovering the institution that owns both. This is the attribution threat that ISAC and NIS2-style sharing programs have not yet confronted: institutional fingerprints aggregate signal from the tool, the analysts, the linters, and the review process simultaneously, and they cannot be removed by individual anonymisation. We frame our temporal result, a 9–20 percentage point accuracy gap between the full history and the 2022–2025 subset, as a premise rather than a conclusion. The effect is consistent in direction and
AC M
Metadata removal is insufficient. Stripping author tags and organisational names does not prevent attribution. Structural patterns in conditions, string definitions, and naming conventions carry author-identifying information that a determined adversary with a modest training corpus can exploit.
families were skipped due to insufficient data. Third, we do not evaluate adversarial robustness: a motivated author could deliberately alter their style to evade attribution. Fourth, our CodeBERT models use single-task training; multitask or few-shot approaches may yield different results. Fifth, the AST features are approximated via regular expressions rather than a formal YARA parser; a full AST extraction might yield stronger structural signals. Sixth, malware family labels were assigned via LLM-assisted categorization rather than ground-truth malware analysis; labeling errors could introduce noise into the Task 4 malware family classification results. However, this does not affect the primary per-family attribution claim (Tier 1 and Tier 2 experiments), which conditions on family membership and is therefore invariant to the absolute correctness of the family taxonomy.
CC
The gap between the two metrics varies systematically across methods and provides diagnostic insight into what type of signal each method relies on. For the malware task, CodeBERT achieves 0.95 weighted F1 and 0.93 macro F1, a narrow gap indicating strong performance even on minority families. In contrast, the AST method shows a much wider gap (0.88 weighted vs. 0.65 macro), revealing that structural features alone fail to discriminate rare families. The n-gram method falls between (0.92 weighted vs. 0.79 macro), confirming that lexical patterns capture some family-specific content that structural features miss. This pattern, where the gap between macro and weighted F1 varies systematically across methods, provides diagnostic insight into what type of signal each method relies on. For author attribution, the pattern is different: all three methods show a narrow macro-weighted gap (CodeBERT: 0.76 vs. 0.75; ngrams: 0.60 vs. 0.57; AST: 0.54 vs. 0.51), indicating that performance is relatively uniform across authors of different prolificacy. This suggests that authorial style, unlike malware content, does not concentrate in a few dominant classes.
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
References
Su
bm
itt
ed
to
20 26
S
AC M
[1] Shamma Alalawi, Saed Alrabaee, Wasif Khan, Issam Al-Azzoni, and Medha Mohan Ambali Parambil. 2026. A Comparative Study on Source Code Attribution Using AI: Datasets, Features, and Techniques. In Security and Privacy in Communication Networks, Saed Alrabaee, Kim-Kwang Raymond Choo, Ernesto Damiani, and Robert H. Deng (Eds.). Springer Nature Switzerland, Cham, 332–353. [2] Bander Alsulami, Edwin Dauber, Richard Harang, Spiros Mancoridis, and Rachel Greenstadt. 2017. Source Code Authorship Attribution Using Long Short-Term Memory Based Networks. In Computer Security – ESORICS 2017, Simon N. Foley, Dieter Gollmann, and Einar Snekkenes (Eds.). Springer International Publishing, Cham, 65–82. [3] Victor Manuel Alvarez. 2014. YARA: The Pattern Matching Swiss Knife for Malware Researchers. https://virustotal.github.io/yara/ Accessed: 2026-03-01. [4] Marco Alvarez-Fidalgo, Eduardo G. Pardo, and Javier Fresno. 2025. CLAVE: A Contrastive Learning Approach for Authorship Verification of Source Code. Information Sciences 690 (2025), 104005. doi:10.1016/j.ins.2024.104005 [5] ANSSI/CERT-FR. 2021. Sandworm Intrusion Set Campaign Targeting Centreon Systems (CERTFR-2021-CTI-005). Technical Report. Agence Nationale de la Sécurité des Systèmes d’Information. https://www.cert.ssi.gouv.fr/cti/CERTFR-2021CTI-005/ [6] Aylin Caliskan-Islam, Richard Harang, Andrew Liu, Arvind Narayanan, Clare Voss, Fabian Yamaguchi, and Rachel Greenstadt. 2015. De-anonymizing Programmers via Code Stylometry. In 24th USENIX Security Symposium (USENIX Security 15). USENIX Association, Washington, D.C., 255–270. https://www.usenix.org/ conference/usenixsecurity15/technical-sessions/presentation/caliskan-islam [7] Soohyeon Choi, Yong Kiam Tan, Mark Huasong Meng, Mohamed Ragab, Soumik Mondal, David Mohaisen, and Khin Mi Mi Aung. 2025. I Can Find You in Seconds! Leveraging Large Language Models for Code Authorship Attribution. arXiv:2501.08165 [cs.SE] https://arxiv.org/abs/2501.08165 [8] Dectot-Le Monnier de Gouville Esteban, Mohammad Hamdaqa, and Moataz Chouchen. 2026. Mining the YARA Ecosystem: From Ad-Hoc Sharing to DataDriven Threat Intelligence. arXiv:2603.14191 [cs.SE] https://arxiv.org/abs/2603. 14191 [9] Atish Kumar Dipongkor, Ziyu Yao, and Kevin Moran. 2025. Reassessing Code Authorship Attribution in the Era of Language Models. arXiv:2506.17120 [cs.SE] https://arxiv.org/abs/2506.17120 [10] ENISA. 2017. Information Sharing and Analysis Centers (ISACs): Cooperative Models. Technical Report. European Union Agency for Cybersecurity. https://www.enisa.europa.eu/publications/information-sharing-andanalysis-center-isacs-cooperative-models [11] European Parliament and Council. 2022. Directive (EU) 2022/2555 on Measures for a High Common Level of Cybersecurity Across the Union (NIS2). Official Journal of the European Union, L 333, pp. 80–152. https://eur-lex.europa.eu/eli/ dir/2022/2555/oj [12] Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A Pre-Trained Model for Programming and Natural Languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for Computational Linguistics, Online, 1536– 1547. doi:10.18653/v1/2020.findings-emnlp.139
[13] Sophia F. Frankel and Krishnendu Ghosh. 2021. Machine Learning Approaches for Authorship Attribution using Source Code Stylometry. In 2021 IEEE International Conference on Big Data (Big Data). IEEE, Virtual Event, 3298–3304. doi:10.1109/ BigData52589.2021.9671332 [14] Georgia Frantzeskou, Efstathios Stamatatos, Stefanos Gritzalis, Carole E. Chaski, and Blake Stephen Howald. 2007. Identifying Authorship by Byte-Level N-Grams: The Source Code Author Profile (SCAP) Method. International Journal of Digital Evidence 6, 1 (2007), 1–18. [15] Samuel Hassine and Julien Richard. 2019. OpenCTI: Open Cyber Threat Intelligence Platform. https://github.com/OpenCTI-Platform/opencti Developed by Luatix, ANSSI, and CERT-EU. [16] Marek Horvath, Emilia Pietrikova, and Diomidis Spinellis. 2026. Bridging Behavioral Biometrics and Source Code Stylometry: A Survey of Programmer Attribution. arXiv:2603.11150 [cs.SE] https://arxiv.org/abs/2603.11150 [17] Christopher Johnson, Lee Badger, David Waltermire, Julie Snyder, and Clem Skorupka. 2016. Guide to Cyber Threat Information Sharing (NIST Special Publication 800-150). Technical Report. National Institute of Standards and Technology. doi:10.6028/NIST.SP.800-150 [18] Vaibhavi Kalgutkar, Ratinder Kaur, Hugo Gonzalez, Natalia Stakhanova, and Alina Matyukhina. 2019. Code Authorship Attribution: Methods and Challenges. ACM Comput. Surv. 52, 1, Article 3 (feb 2019), 36 pages. doi:10.1145/3292577 [19] Kaspersky Lab Global Research & Analysis Team. 2015. Equation Group: Questions and Answers. Technical Report. Kaspersky Lab. https://securelist.com/equationgroup-questions-and-answers/68877/ [20] Florian Roth. 2020. yarGen: YARA Rule Generator. GitHub repository. https: //github.com/Neo23x0/yarGen Version 0.23.2, a generator for YARA rules. [21] Farhan Ullah, Junfeng Wang, Sohail Jabbar, Fadi Al-Turjman, and Mamoun Alazab. 2019. Source Code Authorship Attribution Using Hybrid Approach of Program Dependence Graph and Deep Learning Model. IEEE Access 7 (2019), 141987– 141999. doi:10.1109/ACCESS.2019.2943639 [22] Thomas D. Wagner, Khaled Mahbub, Esther Palomar, and Ali E. Abdallah. 2019. Cyber Threat Intelligence Sharing: Survey and Research Directions. Computers & Security 87 (2019), 101589. doi:10.1016/j.cose.2019.101589 [23] Kai Yao and Marc Juarez. 2026. Smudged Fingerprints: A Systematic Evaluation of the Robustness of AI Image Fingerprints. arXiv:2512.11771 [cs.CV] https: //arxiv.org/abs/2512.11771
CC
magnitude across three independent methods, but it rests on a single repository’s commit history and a four-year window. We therefore offer it as suggestive evidence that fingerprint strength is not static over time, and we call for longer-horizon studies on corpora with richer temporal annotation to test it. The operational implication for the YARA-sharing ecosystem, and for the sector-specific ISACs and national CERTs that anchor it, is that contributor anonymity cannot be treated as a property of the rule header. It is a property of the entire production pipeline that produced the rule, and defending it requires intervention at every stage of that pipeline: the choice of generator, the structure of templates, the enforcement of style guides, and ultimately the willingness to share representations of rules rather than rules themselves. Our findings motivate a research programme on privacy-preserving detection-rule sharing in which attribution resistance is treated as a first-class design objective, alongside detection coverage and operational efficiency. The dataset, the experimental pipeline, and the attribution-verdict framework introduced here are published to support that programme.
A
Open Science
In line with the CCS 2026 Open Science policy, all artefacts required to reproduce the results reported in this paper are released at an anonymous URL for the duration of the review process. The underlying YARA rules are drawn from three publicly available GitHub repositories (Neo23x0/signature-base, reversinglabs-yara-rules, and YARAHQ/yara-forge), whose upstream licences are preserved in the released bundle. Anonymous artefact URL.. https://anonymous.4open.science/r/ anon-yara-attribution-XXXX2 Released artefacts. The repository contains: • Corpus and preprocessing. The unified dataset yara4_with_repo.csv (23,305 rules; columns text, author, malware_family, repo, rule_name) and the time-restricted subset yara2.csv (14,366 rules from 2022–2025, Neo23x0/signature-base) used for author attribution and timed repository attribution. The preprocessing pipeline implementing metadata stripping, blockcomment removal, tag removal, and class-size filtering as described in §5.2 is provided as a reproducible script. • Malware-family label assignment. The two-step LLM-assisted categorisation pipeline used to label Malware family (§5.1), including the fixed taxonomy of 22 families, and the produced rule-to-family mapping.
2 Placeholder; the final anonymous URL is provided in the HotCRP submission form
and will be active from the submission date through the end of the review cycle.
Anonymous YARA Rules Are Not Anonymous
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
commit history and in the meta: fields of the unredacted rules as published by their authors. No private repositories, private feeds, ISAC-restricted material, or TLP-restricted advisories were collected or analysed. Consequently, no individual or organisation is identified in this work beyond what they have already chosen to publish under their own name.
CC
S
20 26
Identification of named individuals and organisations. Our institutionalattribution finding (§6.3) discusses two analysts (Tobias Michalski and Florian Roth) and one organisation (Nextron Systems) by name. Each of these names already appears as an author field in the public rules of Neo23x0/signature-base and in publicly authored technical posts on the organisation’s official blog, both of which we cite. We name them only to make the institutional-coupling argument externally verifiable by the reader; no additional private information is introduced, and no claim is made about either individual beyond their publicly documented affiliation with Nextron Systems and their role in maintaining the Neo23x0/signature-base repository and the yarGen tool. We do not attribute any specific rule to either analyst beyond the attribution already present in the upstream repository. Risk–benefit analysis. The central contribution of this work is a demonstration that metadata removal alone does not preserve contributor anonymity in YARA rule sharing. We considered whether releasing this result could expose specific contributors or organisations to new risk, and concluded that the net effect is defensive.
AC M
• Classifier implementations. (i) the shared CodeBERT classifier module (codebert_common.py), implementing the CodeBERTClassifier and SingleTaskDataset used across all four tasks; (ii) four single-task training drivers (task1_author.py, task2_repo_timed.py, task3_repo_full.py, task4_malware.py); (iii) lexical 𝑛-gram and AST-feature baselines (ast_ngrams_training.py) reproducing the Burrows’ Delta and Caliskan-Islam-style pipelines reported in Table 1. • Confound-controlled experiments. The Tier 1 per-family and Tier 2 per-family × per-repository training scripts (Tables 2– 3), the varying-𝑁 sweep script used for the Hacktool analysis in §6.3, and the confusion-matrix extraction code used for Figure 8. • Figure and table regeneration. Scripts that regenerate every table and figure in the paper (Tables 1–4; Figures 2–8) from the logged training outputs, with data and plotting parameters separated from plotting logic for ease of inspection. • Training logs. Per-task stdout/stderr logs recording hyperparameters, per-epoch metrics, early-stopping events, and validation scores, supporting independent verification of the reported numbers without re-running training. • Reproduction instructions. A README specifying the conda environment (PyTorch 2.6.0+cu118, CUDA 11.8, Transformers, scikit-learn), the command order, the GPU-adaptive execution mode with a CPU fallback, and the expected runtime per task on a single GPU.
itt
ed
to
Model checkpoints. We do not release fine-tuned CodeBERT checkpoints as part of the anonymous artefact bundle. Training is fully deterministic given the released code, the fixed corpus (yara4_with_repo.csv), and the hyperparameters specified in §5.3 (AdamW, learning rate 2 × 10−5 , batch size 16, early stopping with patience 3, seed fixed in the configuration file). Every reported accuracy, macro-F1, and weighted-F1 value in Tables 1–4 can be regenerated end-to-end from the released training drivers. Per-task training logs recording per-epoch metrics, early-stopping events, and final validation scores are included in the bundle for independent verification without re-running training.
B
Su
bm
No human-subjects data. The corpus contains only the source text of YARA detection rules published in public GitHub repositories. No personal data, no private repositories, and no endpoint or sandbox telemetry are used. Author and repository labels originate from the public commit history of the respective repositories and are not the product of any deanonymisation performed by the authors.
Ethical Considerations
This paper studies a privacy risk inherent to the sharing of YARA detection rules. We consider the ethical implications below, following the USENIX Security’26 ethics guidance referenced by the CCS 2026 CFP. Data sources and consent. All rules used in this study are drawn from three public GitHub repositories, each operating under a permissive open-source licence that explicitly allows redistribution and analysis. Author and repository labels are not inferred or reconstructed: they are already present, in plaintext, in the public
• The attribution capability demonstrated here requires only publicly scrapeable rule corpora and standard ML tooling. Any adversary motivated to re-identify rule authors already has access to the same inputs and the same methods. Our paper does not introduce a new capability; it quantifies one that already exists and is unlikely to have gone unexploited. • Defenders, ISACs, national CERTs, and the NIS2-aligned operators of essential services who coordinate detection sharing, currently treat metadata stripping as sufficient protection. Withholding this result would leave that assumption unchallenged. Publishing it allows the community to design countermeasures (§7.5) and to set realistic expectations for contributor anonymity in future sharing frameworks. • The specific institutional-attribution example exposes a coupling between an openly maintained rule generator (yarGen), an openly maintained repository (Neo23x0/signature-base), and an openly branded organisation (Nextron Systems). This coupling is documented on the organisation’s own public channels. We do not believe its discussion creates any reputational, legal, or operational risk beyond what is already inherent in the organisation’s chosen public profile.
Notification. Our findings concern information that Nextron Systems, Mr. Michalski, and Mr. Roth have themselves chosen to publish, on the organisation’s official blog and in the authored rules of the Neo23x0/signature-base repository. The institutional coupling we describe is documented on Nextron Systems’ own public channels and is cited accordingly. We have therefore not treated this work as a vulnerability disclosure, and no embargo or prepublication notification procedure applies. At the same time, we
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY
Anonymous Authors
No IRB/ERB review. The study does not involve human subjects, user data, or interaction with third parties. IRB/ERB review is therefore neither required nor meaningful under our institution’s definition of human-subjects research. We have nonetheless reasoned about the ethics of the work beyond formal institutional compliance, as reflected in this section.
Su
bm
itt
ed
to
AC M
CC
S
Dual-use considerations. The released artefacts (corpus, training code, labels) allow reproduction of our attribution classifiers. We assessed whether release itself poses incremental offensive risk and concluded it does not: the classifiers operate on public data that any adversary can already collect, and the primary audience for the released code is the defender community developing countermeasures. The released pipeline can equally be used to audit the attribution resistance of a candidate rule-sharing protocol, which is an intended use.
LLM-assisted data labelling. Malware-family labels for the yara3 subset were produced by a frontier LLM applied to the rule_name and meta: description fields of each rule (§5.1). These labels are used only as coarse-grained stratification for the per-family experiments; the core attribution claims (Tiers 1 and 2) are invariant to the absolute correctness of the family taxonomy, as they condition on family membership rather than predicting it. We flag this choice explicitly so that reviewers and downstream users can weight the Task 4 malware-classification numbers accordingly.
20 26
recognise that naming individuals in a re-identification paper warrants caution. Accordingly, (i) no claim in the paper goes beyond what is verifiable from the cited public sources, (ii) no specific rule is attributed to an individual beyond the attribution already present in the upstream repository, and (iii) the paper’s framing is squarely defensive, quantifying an attribution surface that ISAC and NIS2-aligned sharing programmes should design against rather than providing a novel offensive capability.