pith. sign in

arxiv: 2411.18384 · v2 · submitted 2024-11-27 · 💻 cs.NI · cs.AI· math.OC

Optimal In-Network Distribution of Learning Functions for a Secure-by-Design Programmable Data Plane of Next-Generation Networks

Mattia Giovanni Spina , Edoardo Scalzo , Floriano De Rango , Francesca Guerriero , Antonio Iera This is my paper

Pith reviewed 2026-05-23 17:34 UTC · model grok-4.3

classification 💻 cs.NI cs.AImath.OC
keywords programmable data planein-network computingintrusion detectionoptimization modelmeta-heuristicnetwork securitymachine learning deployment
0
0 comments X

The pith

A mathematical model distributes machine learning functions for intrusion detection across programmable network devices to maintain security with low added load on each device.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes an optimization model that assigns portions of IDS or IPS learning tasks to different data plane devices so the entire network stays protected while no single device carries excessive extra work. It also introduces a meta-heuristic solver that finds good distributions much faster than solving the full mathematical program. A sympathetic reader would care because this moves security processing inside the network fabric itself rather than relying only on end hosts or external appliances. If the distribution works as claimed, networks gain an autonomous first line of defense that runs continuously on the devices already handling traffic.

Core claim

The central claim is that an optimization model can allocate IDS/IPS machine-learning workloads across programmable data-plane devices to achieve full network coverage while keeping the added processing burden on each device minimal; a meta-heuristic method is shown to produce near-optimal allocations in far less time than the exact solver.

What carries the argument

The mathematical optimization model that decides how to split and place the learning functions for intrusion detection or prevention among the data-plane devices.

If this is right

  • Network devices can run distributed machine-learning detection as their normal forwarding work continues.
  • The meta-heuristic produces usable placements much faster than the exact optimization.
  • The resulting data plane acts as an embedded, autonomous layer of cyber defense.
  • Security functions become part of the programmable fabric rather than added external systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same placement logic could be tested on other in-network machine-learning tasks such as traffic classification or anomaly detection.
  • Real deployments would still need to account for switch memory sizes and exact model footprints that the current model leaves unspecified.
  • Periodic re-optimization might be required when traffic patterns shift, an aspect not examined here.

Load-bearing premise

That an optimal split of the detection workload will deliver complete security coverage while each device's extra processing load stays low, even without detailed hardware limits or changing traffic patterns being modeled.

What would settle it

A simulation or testbed run on a realistic topology that shows either undetected attacks or overloaded devices when the model’s recommended distribution is applied to real attack traffic.

Figures

Figures reproduced from arXiv: 2411.18384 by Antonio Iera, Edoardo Scalzo, Floriano De Rango, Francesca Guerriero, Mattia Giovanni Spina.

Figure 1
Figure 1. Figure 1: Proposed Split-AI In-Network Distribution Strategy. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: From WL-VNFs to Colors domain. The graph edges are weighted to reflect a network connec￾tion characteristic, such as latency or bandwidth. Our objective is to find the optimal deployment of WL-VNFs to ensure comprehensive network security coverage. This approach guarantees pervasive and ubiquitous network protection, aligning with the need for robust cybersecurity measures in the evolving landscape of next… view at source ↗
Figure 3
Figure 3. Figure 3: Average Packet Size for DDoS attack in CIC-DDoS2019. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average Classification Time for Experimental Scenarios: a) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average Throughput for Experimental Scenarios: a) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative distribution of AWDelay for the density classes for each node class. distribution at AWDelay values below about 2%. Specifically, the curves representing ED1, ED3, and ED4 show almost identical behavior, with a high percentage of observed values (90%) having delays below about 1%. The ED2 curve shows a similar trend but with a slightly more gradual increase, indicating greater variability in del… view at source ↗
Figure 7
Figure 7. Figure 7: , where the box plots illustrate the distribution of AWDe￾lay across different densities. In particular, the interquartile ranges expand in sparser networks, showing greater variability in path efficiency due to the limited number of feasible paths that meet the coloring constraints. The box plots also highlight that in more connected networks, such as those with ED4, the AWDelay distribution is more compa… view at source ↗
read the original abstract

The rise of programmable data plane (PDP) and in-network computing (INC) paradigms paves the way for the development of network devices (switches, network interface cards, etc.) capable of performing advanced processing tasks. This allows running various types of algorithms, including machine learning, within the network itself to support user and network services. In particular, this paper delves into the deployment of in-network learning models with the aim of implementing fully distributed intrusion detection systems (IDS) or intrusion prevention systems (IPS). Specifically, a model is proposed for the optimal distribution of the IDS/IPS workload among data plane devices with the aim of ensuring complete network security without excessively burdening the normal operations of the devices. Furthermore, a meta-heuristic approach is proposed to reduce the long computation time required by the exact solution provided by the mathematical model and its performance is evaluated. The analysis conducted and the results obtained demonstrate the enormous potential of the proposed new approach for the creation of intelligent data planes that act effectively and autonomously as the first line of defense against cyber attacks, with minimal additional workload on the network devices involved.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes a mathematical model for optimally distributing IDS/IPS machine-learning workload across programmable data plane (PDP) devices to achieve complete network security with minimal device burden. It introduces a meta-heuristic to approximate the exact solution in reasonable time and evaluates the approach, concluding that it demonstrates enormous potential for intelligent, autonomous data planes as first-line cyber defense.

Significance. The topic addresses a timely intersection of programmable networks and in-network security. A sound model plus practical meta-heuristic could inform secure-by-design PDP architectures. However, the manuscript supplies no equations, validation data, attack-trace results, or hardware-constraint analysis, so the claimed potential cannot be assessed from the provided material.

major comments (3)
  1. [Abstract] Abstract: the central claim that optimal workload distribution yields 'complete network security' with 'minimal additional workload' is asserted without any quantification of detection coverage, false-positive rates, or accuracy after function partitioning. No attack traces, PDP resource measurements (memory, pipeline stages), or dynamic-traffic experiments are referenced.
  2. [Modeling approach] Modeling approach (as described): the premises that (1) ML functions can be partitioned without accuracy loss, (2) PDP hardware permits faithful execution, and (3) a static optimum remains valid under non-stationary traffic are stated but never tested or bounded. These are load-bearing for the 'complete security' guarantee yet receive no supporting derivation or experiment.
  3. [Performance evaluation] Performance evaluation: the meta-heuristic is said to reduce computation time, but no baseline solver times, optimality gaps, or scaling results versus network size are supplied. Without these, the practicality claim and the 'enormous potential' conclusion lack grounding.
minor comments (1)
  1. [Abstract] The abstract contains no equations, parameter definitions, or dataset descriptions, making the contribution difficult to evaluate at the submission stage.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and insightful comments on our manuscript. We address each of the major comments below, providing clarifications and indicating where revisions will be made to strengthen the paper.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that optimal workload distribution yields 'complete network security' with 'minimal additional workload' is asserted without any quantification of detection coverage, false-positive rates, or accuracy after function partitioning. No attack traces, PDP resource measurements (memory, pipeline stages), or dynamic-traffic experiments are referenced.

    Authors: The abstract provides a high-level overview of the paper's contributions and conclusions. The mathematical model ensures complete security coverage by optimizing the placement of IDS/IPS functions across the network, assuming the ML models achieve their intended detection capabilities. The evaluation focuses on the optimization and meta-heuristic performance rather than re-evaluating the ML accuracy or running attack traces, as those are outside the scope. We will revise the abstract to better reflect the assumptions and the nature of the evaluation performed. revision: yes

  2. Referee: [Modeling approach] Modeling approach (as described): the premises that (1) ML functions can be partitioned without accuracy loss, (2) PDP hardware permits faithful execution, and (3) a static optimum remains valid under non-stationary traffic are stated but never tested or bounded. These are load-bearing for the 'complete security' guarantee yet receive no supporting derivation or experiment.

    Authors: These premises are explicitly stated as modeling assumptions in the paper to focus on the workload distribution problem. The paper does not claim to validate the ML partitioning accuracy or hardware execution fidelity, as the contribution is the optimization framework under these standard assumptions from the INC and PDP literature. We agree that a more detailed discussion of the validity of these assumptions would be beneficial and will add a dedicated subsection on model assumptions and limitations in the revised manuscript. revision: partial

  3. Referee: [Performance evaluation] Performance evaluation: the meta-heuristic is said to reduce computation time, but no baseline solver times, optimality gaps, or scaling results versus network size are supplied. Without these, the practicality claim and the 'enormous potential' conclusion lack grounding.

    Authors: The manuscript does include performance evaluation of the meta-heuristic, demonstrating reduced computation times. However, to provide stronger grounding, we will expand the evaluation section to include explicit comparisons with baseline solvers, optimality gap measurements, and scaling results as a function of network size in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity: optimization model and meta-heuristic presented without self-referential reductions

full rationale

The abstract and context describe a proposed mathematical model for optimal IDS/IPS workload distribution among PDP devices plus a meta-heuristic solver, followed by performance evaluation. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The central claim rests on results from the model rather than any derivation that reduces to its own inputs by construction. This matches the default expectation of a non-circular paper whose claims are independent of the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities can be identified from the given text.

pith-pipeline@v0.9.0 · 5751 in / 1081 out tokens · 30958 ms · 2026-05-23T17:34:56.949520+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    Distributing intelligence in 6g programmable data planes for effective in-network intrusion prevention,

    M. G. Spina, F. D. Rango, E. Scalzo, F. Guerriero, and A. Iera, “Distributing intelligence in 6g programmable data planes for effective in-network intrusion prevention,” IEEE Network, pp. 1–1, 2025

    work page 2025

  2. [2]

    Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge,

    Y . Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L. Tang, “Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge,” SIGARCH Comput. Archit. News , vol. 45, no. 1, p. 615–629, apr 2017

    work page 2017

  3. [3]

    Auto-Split: A General Framework of Collaborative Edge- Cloud AI,

    A. Banitalebi-Dehkordi, N. Vedula, J. Pei, F. Xia, L. Wang, and Y . Zhang, “Auto-Split: A General Framework of Collaborative Edge- Cloud AI,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , ser. KDD ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 2543–2553

    work page 2021

  4. [4]

    A Survey on In-Network Computing: Programmable Data Plane and Technology Specific Applications,

    S. Kianpisheh and T. Taleb, “A Survey on In-Network Computing: Programmable Data Plane and Technology Specific Applications,” IEEE Commun. Surv. Tutorials, vol. 25, no. 1, pp. 701–761, Jan. 2023

    work page 2023

  5. [5]

    Native Support of AI Applications in 6G Mobile Networks via an Intelligent User Plane,

    S. Schwarzmann et al. , “Native Support of AI Applications in 6G Mobile Networks via an Intelligent User Plane,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC) , 2024

    work page 2024

  6. [6]

    In-network computing and split-ai in 6g: Enablers and proof-of-concept studies,

    M. Spina et al., “In-network computing and split-ai in 6g: Enablers and proof-of-concept studies,” pp. 1–6, 2024

    work page 2024

  7. [7]

    An intelligent user plane to support in-network computing in 6g networks,

    S. Schwarzmann, R. Trivisonno, S. Lange, T. E. Civelek, D. Corujo, R. Guerzoni, T. Zinner, and T. Mahmoodi, “An intelligent user plane to support in-network computing in 6g networks,” in ICC 2023-IEEE International Conference on Communications , 2023, pp. 1100–1105

    work page 2023

  8. [8]

    Programming the network dataplane,

    C. Kim, “Programming the network dataplane,” ACM SIGCOMM: Florianopolis, Brazil, 2016

    work page 2016

  9. [9]

    In-network Neural Networks

    G. Siracusano and R. Bifulco, “In-network Neural Networks,” arXiv preprint arXiv:1801.05731, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    Can the network be the AI accelerator?

    D. Sanvito, G. Siracusano, and R. Bifulco, “Can the network be the AI accelerator?” in Proceedings of the 2018 Morning Workshop on In- Network Computing, 2018, pp. 20–25

    work page 2018

  11. [11]

    Toward in-network intelligence: Running distributed artificial neural networks in the data plane,

    M. Saquetti, R. Canofre, A. F. Lorenzon, F. D. Rossi, J. R. Azambuja, W. Cordeiro, and M. C. Luizelli, “Toward in-network intelligence: Running distributed artificial neural networks in the data plane,” IEEE Communications Letters, vol. 25, no. 11, pp. 3551–3555, 2021

    work page 2021

  12. [12]

    Taurus: a data plane architecture for per-packet ML,

    T. Swamy, A. Rucker, M. Shahbaz, I. Gaur, and K. Olukotun, “Taurus: a data plane architecture for per-packet ML,” in Proceedings of the 27th ACM International Conference on Architectural Support for Program- ming Languages and Operating Systems , 2022, pp. 1099–1114

    work page 2022

  13. [13]

    Homunculus: Auto-Generating Efficient Data-Plane ML Pipelines for Datacenter Networks,

    T. Swamy et al. , “Homunculus: Auto-Generating Efficient Data-Plane ML Pipelines for Datacenter Networks,” inProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3 , 2023, pp. 329–342

    work page 2023

  14. [14]

    SwitchTree: in-network computing and traffic analyses with Random Forests,

    J.-H. Lee and K. Singh, “SwitchTree: in-network computing and traffic analyses with Random Forests,” Neural Computing and Applications , pp. 1–12, 2020

    work page 2020

  15. [15]

    pForest: In-Network Inference with Random Forests,

    C. Busse-Grawitz et al. , “pForest: In-Network Inference with Random Forests,” arXiv preprint arXiv:1909.05680 , 2019

    work page arXiv 1909

  16. [16]

    Planter: seeding trees within switches,

    C. Zheng and N. Zilberman, “Planter: seeding trees within switches,” in Proceedings of the SIGCOMM’21 Poster and Demo Sessions , 2021, pp. 12–14

    work page 2021

  17. [17]

    Mousika: Enable General In-Network Intelligence in Programmable Switches by Knowledge Distillation,

    G. Xie, Q. Li, Y . Dong, G. Duan, Y . Jiang, and J. Duan, “Mousika: Enable General In-Network Intelligence in Programmable Switches by Knowledge Distillation,” in IEEE INFOCOM 2022-IEEE Conference on Computer Communications. IEEE, 2022, pp. 1938–1947

    work page 2022

  18. [18]

    IIsy: Hybrid In-Network Classification Using Pro- grammable Switches,

    C. Zheng et al. , “IIsy: Hybrid In-Network Classification Using Pro- grammable Switches,” IEEE/ACM Transactions on Networking , 2024

    work page 2024

  19. [19]

    Empowering in-network classification in programmable switches by bi- nary decision tree and knowledge distillation,

    G. Xie, Q. Li, G. Duan, J. Lin, Y . Dong, Y . Jiang, D. Zhao, and Y . Yang, “Empowering in-network classification in programmable switches by bi- nary decision tree and knowledge distillation,” IEEE/ACM Transactions on Networking, vol. 32, no. 1, pp. 382–395, 2024

    work page 2024

  20. [20]

    Distributing Intelligence in 6G Programmable Data Planes for Effective In-Network Deployment of an Active Intrusion Detection System,

    M. G. Spina, F. De Rango, E. Scalzo, F. Guerriero, and A. Iera, “Distributing Intelligence in 6G Programmable Data Planes for Effective In-Network Deployment of an Active Intrusion Detection System,” arXiv, Oct. 2024

    work page 2024

  21. [21]

    Shortest paths with exclusive-disjunction arc pairs conflicts,

    R. Cerulli, F. Guerriero, E. Scalzo, and C. Sorgente, “Shortest paths with exclusive-disjunction arc pairs conflicts,” Computers & Operations Research, vol. 152, p. 106158, 2023

    work page 2023

  22. [22]

    Depth-first search and linear graph algorithms,

    R. Tarjan, “Depth-first search and linear graph algorithms,” SIAM journal on computing , vol. 1, no. 2, pp. 146–160, 1972

    work page 1972

  23. [23]

    On finding the strongly connected components in a directed graph,

    E. Nuutila and E. Soisalon-Soininen, “On finding the strongly connected components in a directed graph,” Information processing letters, vol. 49, no. 1, pp. 9–14, 1994

    work page 1994

  24. [24]

    The multi-parent biased random-key genetic algorithm with implicit path-relinking and its real-world applications,

    C. E. Andrade, R. F. Toso, J. F. Gonc ¸alves, and M. G. Resende, “The multi-parent biased random-key genetic algorithm with implicit path-relinking and its real-world applications,” European Journal of Operational Research, vol. 289, no. 1, pp. 17–30, 2021

    work page 2021

  25. [25]

    The irace package: Iterated racing for automatic algorithm configuration,

    M. L ´opez-Ib´a˜nez, J. Dubois-Lacoste, L. P ´erez C´aceres, M. Birattari, and T. St ¨utzle, “The irace package: Iterated racing for automatic algorithm configuration,” Operations Research Perspectives , vol. 3, pp. 43–58,

    work page

  26. [26]

    Available: https://doi.org/10.1016/j.orp.2016.09.002

    [Online]. Available: https://doi.org/10.1016/j.orp.2016.09.002

    work page doi:10.1016/j.orp.2016.09.002 2016

  27. [27]

    De- veloping Realistic Distributed Denial of Service (DDoS) Attack Dataset and Taxonomy,

    I. Sharafaldin, A. H. Lashkari, S. Hakak, and A. A. Ghorbani, “De- veloping Realistic Distributed Denial of Service (DDoS) Attack Dataset and Taxonomy,” in2019 International Carnahan Conference on Security Technology (ICCST). IEEE, 2019, pp. 01–03

    work page 2019

  28. [28]

    Timely Detection and Mitigation of Stealthy DDoS Attacks Via IoT Networks,

    K. Doshi, Y . Yilmaz, and S. Uludag, “Timely Detection and Mitigation of Stealthy DDoS Attacks Via IoT Networks,” IEEE Trans. Dependable Secure Comput., vol. 18, no. 5, pp. 2164–2176, Jan. 2021

    work page 2021

  29. [29]

    A realistic cyber defense dataset (cse-cic-ids2018),

    “A realistic cyber defense dataset (cse-cic-ids2018),” accessed: 2023- 04-04. [Online]. Available: https://registry.opendata.aws/cse-cic-ids2018

    work page 2023

  30. [30]

    A tool for the generation of realistic network workload for emerging networking scenarios,

    A. Botta, A. Dainotti, and A. Pescap `e, “A tool for the generation of realistic network workload for emerging networking scenarios,” Computer Networks, vol. 56, no. 15, pp. 3531–3547, 2012

    work page 2012

  31. [31]

    Detecting and Mitigating Botnet Attacks in Software-Defined Networks Using Deep Learning Techniques,

    M. W. Nadeem, H. G. Goh, Y . Aun, and V . Ponnusamy, “Detecting and Mitigating Botnet Attacks in Software-Defined Networks Using Deep Learning Techniques,” IEEE Access, vol. 11, pp. 49 153–49 171, 2023

    work page 2023

  32. [32]

    TLC UNICAL In-network-Distributed-IDS,

    “TLC UNICAL In-network-Distributed-IDS,” Nov. 2024, [Online; accessed 18. Nov. 2024]. [Online]. Available: https://github.com/ mattiagiovanni/TLC UNICAL In-network-Distributed-IDS

    work page 2024

  33. [33]

    behavioral-model,

    “behavioral-model,” Sep. 2024, [Online; accessed 5. Sep. 2024]. [Online]. Available: https://github.com/p4lang/behavioral-model Mattia Giovanni Spina is a PhD student at the University of Calabria (Italy). His research interest is in the area of security in future generation networks and distributed AI in-network architectures. Floriano De Rango is full p...

    work page 2024