Code and Lab
The Federated DDoS Research repository holds the reusable code. Its original code and documentation use the MIT license. The repository currently provides dataset audits, bounded SGD/Adam training checks with FedAvg, portable engineering bundles, the frozen controlled study design v2.1, and scoped Lean proofs with 196 checked theorems in six modules. These components remain research prototypes.
Latest Documentation Checkpoint
November 15, 2026, pre-registered controlled laboratory study, design v2.1. A new controlled laboratory study was designed and frozen before any build, traffic or data. It asks how packet sampling at a provider’s border routers changes what a router-side detector can see during flood-shaped DDoS attacks (RQ2), and how detection at a recipient point of presence changes when a second, contributing point of presence collaborates (RQ1 and RQ4). The study is pre-registered and separate from the earlier controlled track, which stopped at its pre-registered separability gate. That earlier result stays reported unchanged. No laboratory component of the new study has been built or started, and no traffic has been generated. Building the laboratory, qualifying it and acquiring data each need a separate release by the researcher.
The frozen design, topology, propositions and figure are published byte-identical to the frozen files, and their README states the status and limits of the design. The Lean project now holds 196 checked theorems in six modules, including 124 for the design propositions P1 to P12. The checker passes with the standard axioms only (propext, Classical.choice and Quot.sound). The formal-verification README states the limits of that check. For example, P4 is proved for an integer size only, P6 (d), P7 (b) and P9 are proved on finite sample spaces, P10 has its covariance identities only, and the keyed packet selector is treated as independent selection without proof. These are mathematical statements about the design, not measurements. The frozen design establishes what will be measured and how results will be judged. It does not establish detection performance, the effect of sampling, the benefit of collaboration or the behavior of a real provider network.
The next planned task is to write the offline laboratory build artifacts from the frozen design, using benign test fixtures only, and to have them independently audited. The researcher will then be asked for a go-ahead before the qualification and peak-load dry run. The laboratory virtual machine stays stopped until that release.
Earlier November 14, 2026, conditional source reconstruction and fit preparation. A completed reconstruction of previously used development evidence has passed an independent saved-output audit. Sampling and timeout views were replayed against retained ingress evidence, preserving the existing training and development assignments and common evaluation opportunities. Missing sampled observations remain coverage outcomes, not inferred classification decisions. This qualification is limited to the historical strict-conditioned cohort. It does not admit a wider source population, independent episodes, natural provider or customer roles, or untouched calibration and final-evaluation sets. New study fitting and scoring remain unrun.
Development training-controller checks and a separate statistical federated-method plan have also been audited. Actual-source fitting still depends on trusted readers, preserved training context and a bounded release. The methods retain different supervision and weighting, so these checks establish neither matched training work nor detection efficacy. The private English draft now records these boundaries; human reading and acceptance remain open.
Earlier November 14, 2026, ingress functional boundary and source screening. An independently audited finite native check exercised invented ingress packets, forwarding observations and flow-meter replay under the declared observation policy. The saved evidence also supports cleanup of the resources created for that check. This check included no learning or inference and does not establish detection efficacy, endpoint acceptance, deployment capacity, latency or source-data admission. A separate annotation-availability screen has completed; packet association, source eligibility, data-role assignment and fitting remain open gates.
The research plan retains four required questions and a conditional Sellier-inspired fifth question. The required work covers matched independent local, pooled-central and FedAvg learning with a separate statistical federated-system comparison; paired packet-sampling and timeout changes; router deployment with separate forwarding, inference and resource costs; and contributor-policy compatibility versus complementary DDoS behavior. The finite functional checks do not complete these evaluations.
The earlier project status note records the earlier development checkpoint.
Earlier November 14, 2026, development evidence checkpoint. A frozen FLNET comparison has been executed under recorded development settings. Reused observations, different transforms and optimizer choices limit causal and final-evaluation claims. A separate common-feature pilot used packet and octet counts from the same flow meter for independent local, pooled-central and federated training. Under its fixed decision rule, the learned arms and the count baseline produced no positive alerts. This unfavorable finding remains part of the record; it does not establish a causal advantage for federated learning.
An earlier router-local component check exercised a source-trained model on synthetic probes inside the border-router process, together with forwarding, flow-meter and collector-input parity. It establishes that finite component path, not detection efficacy or networked federated training. A distinct two-PoP native toy then exercised invented federated updates, model delivery and loading, and toy-model inference. Its benign toy records were flagged as positive. These false positives do not establish detection efficacy, and the invented updates do not establish source-backed federated training. Neither check establishes operational DDoS detection, scalability, proprietary-router compatibility or compatibility with the separate CSV feature profile.
Paired observation-policy experiments, representative workloads, source validity and broader evaluation remain open. Human manuscript reading and acceptance remain separate gates, followed by French translation and proofreading. The next work is to close the specific evidence and reporting gaps without changing the recorded unfavorable finding.
The exact common-feature and native implementations remain outside the public code release. Reusable packaging is pending a bounded public-source extraction and clean-checkout check that excludes models, data and private execution records. Existing public code links retain their historical scope.
November 12, 2026, frozen CSV scorer publication. The published score-only CSV interface provides export and loading interfaces for neural and AF-BKM-family model snapshots, owner-local TRAIN transform recovery and summary-only merging. Local and hosted checks of the integrated repository each discovered 328 unit tests, with 327 executed and one optional TShark test skipped. The engineering self-test and fabricated workflows passed. These checks use fabricated inputs; they do not establish detection accuracy, calibration or compatibility with live traffic.
The raw-data work supports only a narrow offline connection and counter reference from one inspected capture. Capture completeness, historical analyzer settings and independent attack episodes remain unresolved, so it has not been accepted as a controlled sampling/timeout benchmark. Final protocol selection, functional verification, and measured overlap and capacity checks remain open for the approved lab architecture and IP plan.
The original code commits and unchanged dependency lock remain recorded. This publication adds no acquired-data experiment and contains no fitted models, private execution receipts or raw records. Verification of reporting-threshold derivation and identity-linked decisions remains separate work.
Earlier November 12, 2026, loader checkpoint. The checked role-aware loader reproduced the development records and their existing training, validation and calibration assignments. The required suite passed 77 unit tests, its engineering self-test and synthetic workflow. Two bounded independent code critiques found no consequential implementation defect. A separate feature-value and multiplicity audit remains pending before fitting. No learned preprocessing, model fitting, calibration or model scoring was performed.
This preparation check does not establish training capacity or deployment performance. The candidate retains source observers and reserved-test boundaries; stored records do not establish independent customers or a common original sampling/timeout policy.
Grouped development-role preparation
November 12, 2026. The versioned preparation guide describes a provisional FLNET training, validation and calibration allocation. It preserves source observer IDs and assigns the same stored key to one role wherever it appears. The candidate accounts for the inspected development records and stored keys. It does not open reserved test sources, fit a model or freeze an evaluation protocol. Source files, configuration, code hashes, counts, timing and failure boundaries are recorded.
The earlier role-preparation checks comprised 56 unit tests, the engineering self-test and the 225-record synthetic workflow with deliberately unreadable reserved inputs. The bounded xAI code and aggregate-count critique found no blocking implementation defect for retaining the candidate. A repeated preparation reproduced its assignment digest and all role counts. No scientific protocol approval is inferred. These checks establish the reported preparation behavior. They do not establish independent sessions, operational customers, future-traffic validity or detection performance. Some source observers lack attacks; the final feature and calibration choices remain open. The earlier engineering versions below retain their own code and results.
The current proposed deployment places software inference inside Internet border routers, with collection and local training at PoPs. A finite native-container check has now exercised that inference placement with synthetic probes and a trained model; broader operational behavior remains unverified. The November 4 PoP-scoring assessment below is design history.
Earlier provider observation and learning assessment
November 4, 2026, historical assessment. The earlier provider design uses provider-owned collection and scoring at points of presence (PoPs), with central coordination. These PoPs are candidate learning participants, whereas customers remain reporting and evaluation groups. No detector appliances are assumed on customer premises. The planned comparison covers independent local, permitted pooled-central and federated training. It keeps the same observation view, feature definitions and PoP inference placement. Pooled and federated training use the same permitted fitting evidence. Independent local training retains its local evidence. The comparison remains unrun. Three additional research papers and two RFCs have been acquired, and selected full-paper FlowSentry sections now replace the earlier abstract-only check. No EVE topology or scientific code was changed. The public-source architecture critiques did not receive or review the revised private Chapter 1.
The current working collection contains 103 parent references, including 98 stored paper or book PDFs and five supporting records. The public reference selection now has 23 entries. The project page records the latest checked manuscript status.
The public assessment note provides the same bounded architecture discussion below. It records source roles and reading limits; the comparison has not been run.
That earlier provisional provider-operated design places collection and scoring at provider points of presence (PoPs), with central coordination. It does not assume detector appliances on customer premises. Internet-border ingress is the primary observation candidate for externally offered load. Customer-facing provider-edge (PE) views remain complementary or a separate observation comparison.
Observation, flow collection, training, inference and mitigation remain separate decisions. The proposed training comparison fixes the observation view, feature definitions and PoP inference placement. Pooled and federated training use the same permitted fitting evidence. Independent local training retains its local evidence. Reporting by customer does not, by itself, define the training participants. The working design instead treats provider PoPs as the candidate learning units. These are proposed choices, not a selected final protocol or a completed experiment.
Observation does not determine model placement
RFC 5470, Sections 2, 4 and 5, distinguishes observation, metering, export and collection. A collector can receive records from multiple exporters. Central analytics therefore need not mean a single physical observation point or an inline packet bottleneck. The RFC is an informational architecture document, updated by RFC 6183 on mediation; it does not prove that a particular exporter supports the required egress measurement or fields.
A customer-facing PE view can identify a customer’s traffic when interface and routing context are retained. However, its meaning depends on the actual metering stage. An egress record must not be treated automatically as proof of post-queue delivery to the CE. Earlier congestion or drops can also limit what is observed. Internet-border measurements offer a different view of externally offered load. Earlier observation does not by itself prove an earlier alert. These are information-limit checks for the proposed comparison, not measured outcomes.
The NETSCOUT design example, pp. 1–3, describes provider-wide analytics using flow, BGP and SNMP data, customer/service monitoring, and separate upstream mitigation actions. This 2020 vendor document illustrates an existing deployment pattern. It is not an independent performance comparison or a choice of software for this lab.
Compare centralized and federated learning fairly
A centrally trained model can run inference at provider locations. Local inference alone therefore does not establish a need for federated learning. Pourahmadi et al. (2023) already compare cloud-centralized, edge-silo and federated approaches. Section 5.6 and Figure 5 of the inspected author manuscript describe a centrally trained model used for edge-server detection. Their results are specific to that setup.
| Candidate | Check for the unrun comparison |
|---|---|
| Internet-border observations with central analysis | Cover all relevant borders and establish which externally offered traffic is visible. |
| Customer-facing PE views | Check customer context, metering stage and missing information before treating these as complementary or alternative inputs. |
| Fixed observation view for local-only, pooled-central or federated training | Compare quality and cost at the same PoP inference placement. Pooled and FL fits share permitted fitting evidence; the independent local fit retains local evidence. |
| Combined border and PE observations | Establish provenance and handle overlapping records before assessing additional coverage and cost. |
Pooling depends on permission, governance and operational constraints, not only on common provider ownership. Federation should be justified against the actual objectives. Keeping records local does not itself provide a formal privacy guarantee (Kairouz et al., 2021). Transmission savings also require measurement. The comparison should account for serialized telemetry, model exchange and update cadence over the declared training lifecycle.
Keep the observation view and feature definitions fixed when comparing training methods. Changing the observation location at the same time would mix information and learning effects. Separate checks must establish exporter behavior, sampling, record loss, feature compatibility, inference placement and end-to-end alert timing before conclusions are drawn.
Existing precedents
Published precedents already cover parts of this architecture. RFC 9387 (2023), Sections 3.1 and 3.3, describes forwarding nodes, IPFIX flow collectors, orchestration and machine-learning feedback. Selected sections of the acquired FlowSentry paper (He et al., 2025) describe a central detector receiving records from multiple boundary routers, Louvain router communities and community-specific vector autoregression. Sections 5.2–5.3 describe those methods; Sections 6.1 and 6.3 describe the central setup and a router-scale experiment that holds total flow count and each flow’s traffic scale constant. This targeted section reading supersedes the earlier abstract-only check. It does not establish a complete paper appraisal or reproduction. Multi-point collection, central analysis or router grouping alone therefore does not establish novelty.
The follow-up also acquired Nagy et al.’s SCLDDOS2024 dataset paper (2025) and Barrett et al.’s FADES paper (2026). Their titles and primary records identify further reading on cross-domain training and streaming intrusion detection. Acquisition does not establish a full appraisal or add another experimental factor to the study.
Assessment status
Three bounded xAI Grok critique calls on public architecture concepts and source summaries informed this assessment. The tool reported grok-4.7-build. These critiques did not receive or review the revised private Chapter 1. The writer accepted the central baseline and the separation of design choices, then qualified pooling permission, metering-stage semantics and alert-timing claims. The review used supplied public-source packets and was instructed not to query RAG or other files. Web access was disabled. No independent full-source retrieval is claimed. The critiques do not validate the lab, select the final protocol or certify novelty.
Scalability remains an evaluation goal. Any proposed mechanism needs measured quality/cost improvements and further full-text appraisal of the closest precedents before a novelty claim. No EVE topology was rebuilt, no new training was run and no empirical comparison is reported here. The earlier version-1 lab drawing remains design history. The reviewed scientific code remains unchanged. The Research hub records continuing project status; private drafts and detailed study decisions remain outside this public note.
Assessment references
- Barrett, S., Dorai, G., Li, L., & Rajaganapathy, S. (2026). FADES: Adaptive drift estimation via conformal signals for streaming intrusion detection. Electronics, 15(10), 2114. DOI record
- Hayashi, Y., Chen, M., & Su, L. (2023). Use cases for DDoS open threat signaling (DOTS) telemetry (RFC 9387). RFC Editor. Retrieved from https://www.rfc-editor.org/rfc/rfc9387.html
- He, X., Xie, X., Wang, X., Zhang, L., Xie, K., Chen, L., & Cui, Y. (2025). FlowSentry: Accelerating NetFlow-based DDoS detection. In Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security (pp. 1589–1603). Association for Computing Machinery. DOI record
- Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A. N., … Zhao, S. (2021). Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1–2), 1–210. DOI record
- Nagy, B., Orosz, P., Bozsó, K., & Skopkó, T. (2025). Enhancing DDoS detection: A novel real-world dataset for generalizable cross-domain training. In 2025 21st International Conference on Network and Service Management (CNSM). IFIP. Official paper
- NETSCOUT. (2020). Arbor Sightline: How it works and where it lives in the network. Retrieved from Official PDF
- Pourahmadi, V., Alameddine, H. A., Salahuddin, M. A., & Boutaba, R. (2023). Spotting anomalies at the edge: Outlier exposure-based cross-silo federated learning for DDoS detection. IEEE Transactions on Dependable and Secure Computing, 20(5), 4002–4015. DOI record
- Sadasivan, G., Brownlee, N., Claise, B., & Quittek, J. (2009). Architecture for IP flow information export (RFC 5470). RFC Editor. Retrieved from https://www.rfc-editor.org/rfc/rfc5470.html
November 3, 2026. The English manuscript now establishes the DDoS problem and relevant learning concepts before its technical study detail. The checked English manuscript remains a private working draft. The retained reading bank contains 95 works. The working Zotero collection has 98 parent references because it also includes three supporting engineering and measurement records. The public bibliography was a separate selection of 20 entries at that earlier checkpoint. Required source-data checks, the final learner and protocol, and the novelty assessment remain open. Chapter 1 awaits human review. No new training run, numerical QiML engine or final detection result is part of this milestone. Historical scientific code and engineering measurements remain unchanged.
Foundational reference selection
The foundational-reference note records the focused selection procedure, eight scholarly additions, three supporting records, primary publication links and the targeted reading limits. The public reference library contained 20 APA 6 entries at that checkpoint. The November 4 follow-up adds three papers, bringing the current selection to 23. Citation influence was a secondary discovery aid, not a quality ranking or evidence of novelty.
Separate Antigravity and xAI source-selection critiques returned no blocking findings. Antigravity explicitly selected Claude Opus 4.6 Thinking without an exposed serving identifier; xAI reported grok-4.7-build. Their input contained public-source summaries and explanatory requirements, not revised chapters or full papers. A Grok output-format retry is recorded; model agreement does not certify novelty or validity. The reviewers received bibliography, bounded primary evidence and generic requirements. They did not receive chapters or raw data, independently inspect the whole library, execute experiments or validate the revised manuscript. The writer checked their advice against source and artifact records.
An earlier independent introduction review is complete. Google Gemini and xAI Grok reviewed the same frozen input, and the writer recorded the outcome. At that earlier checkpoint, Chapter 1 remained unchanged. Human reading and any resulting revision, source verification, protocol validation, and empirical evaluation remain separate gates. This checkpoint adds no new training run or final result and does not certify novelty.
A file-format compatibility issue in the private introduction review copy was corrected and checked in desktop Word. Human reading remains pending. No research findings or results changed.
The presentation revision adds original vector figures and native editable evidence tables to the private draft. The figure sources remain editable or reproducible. Each has separate source and output checks. This does not change the pilot code or establish a new numerical training result.
Research assistance and validation
The research-assistance workflow records the purpose and scope of large language model (LLM) assistance, the author’s responsibility, and the status of separate model critiques. Completed reviews retain their recorded model identities; a planned review does not rename earlier work. Model agreement is not scientific validation. Formal mathematics, numerical checks, software checks, controlled experiments, and human review address different questions. This documentation adds no new proof, training run, or final result.
Current Engineering Checkpoint
November 1, 2026. The original benign-only engineering pilot completed four target cases, following 35 unit tests, the existing engineering self-test, and a synthetic integration check. The run used previously inspected development observations and preserved the publisher test boundary. It checks the software path; its outputs are not confirmatory detection results.
Grok subsequently requested revisions. Version 2 addresses confirmed issues in label and control validation, row accounting, failure reporting, Parquet engine selection, and donor seeds. The revised code passed 43 unit tests, the engineering self-test, and synthetic integration. These checks used generated fixtures. The original development run and its executed-source hash remain unchanged; version 2 has not been rerun on real data.
The revised pilot is published at commit 102293e69f2f9e7cdcac58cb12ffe8d7d0165833. Its source, regression tests, and method notes identify the reviewed version. Separate Gemini and Grok reviews both returned PASS on the supplied source and synthetic evidence. The software remains a research prototype; these reviews do not establish detection performance, independent episodes, or novelty. Exact executed-source hashes, configuration, dependency lock, and run artifacts are preserved in the private research record.
Review execution check
Earlier Grok requests reached their local deadlines without final answers. A later authorized retry completed and requested revisions. The revised-code reviews then returned PASS on the supplied bounded packet. These are review outcomes, not measurements of model thinking time or a new real-data training run.
Progress records
The private progress record retains completed preparation and pending human reading separately. Recorded workflow checks do not establish chapter readiness or overall research completion. English drafting and revisions come first, followed by French translation and proofreading.
Detailed experiment and timing records remain private. No new training was performed during this literature and manuscript checkpoint, and the historical engineering checks do not establish an optimizer ranking or predict the full study runtime.
Code Versions and Results
The historical public project status note from November 3, 2026 records that earlier working-draft checkpoint and links back to this research hub. The critical literature review workflow describes how source claims, reading limits, and counterevidence are recorded.
The figure and table workflow records the checked conversion routes and their limits. The new commit contains documentation; the executed and reviewed training code remains at its recorded earlier version.
For a reproducible result, record the full commit SHA, configuration, dependency lock, input provenance, and run identifier. A link to the current branch is useful for browsing; a commit permalink identifies the code used in a particular experiment. The repository contains citation metadata and a methodology guide.
Container Deployment
The finite native-container check above uses a separate controlled local route. The following EVE distribution plan remains historical preparation; it is not the route verified by that check. Public images and reusable packaging remain pending.
Container packaging is the proposed next distribution step. Source and build instructions will remain in GitHub, and tested images can be published in GitHub Container Registry. The registry supports public OCI images and pulls by immutable digest. No project container image has been published or tested in EVE at this stage. See the GitHub registry documentation.
A container packages an application and its dependencies. A runtime still has to start it on the destination machine. Docker separates the client that requests a pull or run from the daemon that manages images and containers. See the Docker architecture documentation.
The proposed lab route uses EVE Professional Docker nodes directly. Its cookbook documents pulling a simple third-party image with dc pull <image> on the EVE host and then selecting that image in a topology. This is documented command syntax, not a command tested with this project’s image. The dc command targets EVE’s Docker service at tcp://127.0.0.1:4243. This route does not require a separate Ubuntu guest. See section 14.6, pp. 218–219 of the EVE Professional cookbook.
The initial environment inventory recorded EVE Pro 6.5.0-27. The cited cookbook is version 7.4 and describes third-party container support dating from Pro 2.0.6-52. The feature is documented, but startup and networking with this project’s image still require checks on the installed version.
The cookbook describes running this pull from a host terminal. The website account does not itself establish host-terminal access. Acquiring the project’s own image still needs verification on the installed version. Before collecting research traffic, test the image’s startup behavior, lab interfaces, console, capture permissions, and persistence. Multi-container stacks have a separate EVE integration procedure; they should not be started beside the host’s own services without that integration.
The initial package should run the existing offline engineering checks. Packet capture, traffic generation, model inference, and a networked FL coordinator have different roles and require their own implementation and validation. Packaging the offline tools does not establish a live detector or a completed FL lab.
Release Record
| Material | Current state |
|---|---|
| Research source | Public GitHub repository, including the reviewed version-2 pilot |
| Proposed lab figure | Available on the project page |
| Container image | Planned; no pull command is advertised yet |
| Networked FL lab | Separate two-PoP invented-update toy and ingress functional check executed; source-backed training, operational efficacy and cost evaluation remain open |
| Controlled study design v2.1 | Frozen and published unchanged on November 15, 2026. The planned laboratory is one isolated virtual machine with no route out, and all traffic sources are scripted processes on one computer, so it cannot show distributed attack impact or Internet-scale behavior. No generator code, packet bytes or attack fixtures are published. Nothing is built, and no traffic has been generated. |
| Lean checks | 196 theorems in six modules, 124 of them for design propositions P1 to P12, with standard axioms only; limits are in the README |
| Confirmatory results | Not yet available |
The website will link each published release to its code version, image digest when applicable, and supporting records. Experimental datasets, private drafts, credentials, and correspondence do not belong in the public image.