Pith. sign in

REVIEW 3 major objections 4 minor 82 references

CyberSentinel: Efficient Anomaly Detection in Programmable Switch using Knowledge Distillation

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read CyberSentinel claims the first accurate line-speed anomaly detector that runs entirely in the programmable switch data plane, using an Isolation Forest distilled from autoencoder reconstruction errors to match control-plane accuracy at…

desk verdict Real Tofino prototype with a genuinely new unsupervised knowledge-distillation scheme, but the headline 'similar detection performance' only holds for the deployed 8-feature version at 40–200x higher false-positive rates; the 21-feature version that matches at the same FPR is simulated, not on the switch. read the letter →

arxiv 2412.16693 v1 pith:T2EMGPJD submitted 2024-12-21 cs.CR cs.NI

classification cs.CRcs.NI
keywords anomalydetectionprogrammableswitchP4dataplaneknowledgedistillationIsolationForestautoencoderIoTsecurityline-speed
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CyberSentinel claims that accurate, unsupervised anomaly detection can run entirely inside a programmable switch's data plane, at line speed, without a control-plane model in the detection path. The paper's method distills the learned knowledge of an ensemble of autoencoders into an Isolation Forest, then converts the distilled forest into whitelist rules that P4 match-action tables can evaluate. On a 40 Gbps testbed, the deployed prototype matches the detection performance of the best control-plane-assisted system while increasing packet-processing throughput by 66.47% and cutting average per-packet latency by 50%. If correct, this removes the control plane as the throughput bottleneck, making autoencoder-grade detection scalable to high-rate IoT traffic.

What carries the argument

The load-bearing object is the distilled Isolation Forest: an unsupervised tree ensemble that isolates anomalies by short path lengths, whose leaves are augmented with expected reconstruction errors from an ensemble of autoencoders and labelled benign or malicious (Eqs. 2-3). A second piece, the burst-to-flow mapping module, converts variable-length sequences of burst-level features into the fixed-length flow vectors that popular autoencoders expect, which is what makes the distillation possible when teacher and student are trained on different feature granularities. On the hardware side, the key mechanism is a P4 resubmission-free design: register read/update/reset actions are arranged atomically within the switch's action unit so that no packet is recirculated for feature maintenance, which is the main source of the throughput and latency gains.

What would settle it

Evaluate the burst-to-flow mapping module in isolation against an oracle: feed the same flows' true flow-level features directly to the autoencoder, compare those reconstruction errors to the ones obtained through the learned burst-to-flow mapping, and compute TPR/TNR at the paper's operating points. Alternatively, re-run the end-to-end comparison with the learned mapping replaced by the oracle flow features; if detection performance moves beyond a small margin, the mapping, not the distillation, carries the result.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is a way to embed the reconstruction-error signal of a trained autoencoder ensemble into the leaf nodes of an Isolation Forest without altering the forest's structure. During training, each leaf is labelled by the weighted majority of the base autoencoders' per-leaf average reconstruction errors, so inference is just a tree traversal and a majority vote. The labelled forest is then compressed into a small set of hypercubes, and hypercubes labelled benign become whitelist rules installed in the switch. The paper argues that this distilled iForest retains the high true-positive rate of the original iForest and the low false-positive rate of the autoencoder, delivering the first fully data-plane, line-speed detector with accuracy comparable to a control-plane autoencoder.

Load-bearing premise

The load-bearing premise is that the deferred burst-to-flow mapping module faithfully preserves the autoencoder's reconstruction-error signal when it converts variable-length burst statistics into fixed-length flow vectors, so if that conversion distorts the signal, the distilled forest's low-false-positive advantage erodes.

Editorial extensions

If this is right

  • Anomaly detection no longer has to stop or slow down for the control plane; the switch itself can mark malicious bursts and trigger blacklist insertion.
  • The detection scheme remains unsupervised: it needs only benign traffic for training, so it can flag unseen attacks without labels.
  • Because the deployed artifact is a small set of whitelist rules, the memory footprint stays low (about 9.86% SRAM, 2.85% TCAM in the reported prototype), leaving room for other in-network functions.
  • The same distillation recipe should transfer to other unsupervised teachers and other switch targets, since the teacher is only used offline.
  • Model updates can be continuous: normal bursts mirrored to the control plane retrain the distilled forest and refresh the whitelist rules.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the burst-to-flow mapping is as lossless as the end-to-end numbers imply, the same distillation should work for any teacher whose input features are flow-level, suggesting a general recipe for porting deep unsupervised detectors into switches.
  • Editorial inference: the reported 66.47% throughput gain bundles two separate effects — the distillation (which removes control-plane round trips) and the resubmission-free P4 design. A controlled ablation that isolates these two would tell operators how much of the gain transfers to a switch that already avoids resubmissions.
  • Editorial inference: the consistency metric C is computed against the teacher's own labels, so high C does not by itself guarantee better end-task detection; a direct comparison on TPR/TNR at matched FPR is the more operationally meaningful figure.
  • Editorial inference: if the whitelist-rule compression merges adjacent same-label hypercubes, the rule set could become coarser as traffic shifts; periodic online retraining (which the paper implements) is what keeps the rules aligned with drift.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. CyberSentinel proposes a fully data-plane anomaly detection system for programmable switches. The authors train an ensemble of autoencoders (Magnifier) and distill its reconstruction-error knowledge into an Isolation Forest, then convert the distilled iForest into whitelist rules that can be installed on an Intel Tofino switch. The system is evaluated on a 40 Gbps testbed across multiple IoT attack datasets, reporting 39.6 Gbps throughput versus 23.79 Gbps for HorusEye, roughly 50% lower per-packet latency, and detection metrics claimed to be similar to HorusEye. A 21-feature variant is evaluated in simulation because the deployed 8-feature version cannot fit the switch memory.

Significance. If fully supported, this would be a significant contribution: it would be the first demonstration that autoencoder-grade unsupervised anomaly detection can run inline in the switch data plane without a control-plane bottleneck. The paper's real strengths are the actual Tofino deployment, the P4-level description of a no-resubmission data-path design, the extensive attack-dataset evaluation, and the supplementary algorithms and P4 code. However, the headline claim of "similar detection performance" is not currently established at comparable false-positive rates for the deployed hardware, and the burst-to-flow mapping module that the distillation depends on is not evaluated standalone. These issues are load-bearing for the central contribution and must be fixed before the claim can be accepted.

major comments (3)
  1. [§VII-D and Table III; Abstract]
  2. [Supplementary §XIV, Algorithm 4]
  3. [Supplementary §XIV.G and Algorithm 4]
minor comments (4)
  1. [§VII-D, text after Table III]
  2. [Eq. (5) and Fig. 15]
  3. [Supplementary §XI, Theorem 3]
  4. [General presentation]

Circularity Check

1 steps flagged · score 4.0 of 10

Distilled-iForest leaf labels are thresholded autoencoder errors, so the reported consistency/fidelity metric is self-referential; core throughput and ground-truth detection results remain independent.

  1. self definitional [Section IV-A1, Eqs. (2), (3), (5); echoed in Section VII-D and Section XIII-A]
    "We embed the reconstruction errors (REs) of the r autoencoders in each of the leaf node ... REleaf = (1/(k+|Xleaf|) * sum_{x in Xleaf union Aleaf} REu(x))_{u=1}^{r} (2) ... labelleaf = 1 { sum_{i=1}^{r} w_i * l_i > 0.5 } (3) ... Consistency of knowledge distillation ... C = 1/N * sum_{i=1}^{N} 1 { iForest_distilled(x_i) = Autoencoder(x_i) } (5)"

    The distilled iForest's per-leaf label is the thresholded mean of the autoencoder's reconstruction-error decisions (Eqs. 2-3). Eq. (5) then validates the distilled model by counting agreement with Autoencoder(x_i), the same function that produced the leaf labels. High C therefore measures how finely the leaves reproduce the teacher on the evaluation set; it is a quantization/approximation-fidelity check, not independent evidence of detection capability. The statement 'Since Magnifier's knowledge is transferred in CyberSentinel, we can argue superior attack detection performance similar to that of HorusEye' is likewise an argument from construction.

full rationale

The central claim—accurate anomaly detection entirely in the switch data plane—rests on two pillars. Pillar 1 (line-speed deployment) is fully independent: it is established by the Tofino implementation (Sections IV-C and VII-C) with measured throughput 39.59 Gbps vs 31.02 Gbps and latency 532.8 ns vs 1065.6 ns versus Gulliver Tunnel, plus resource-utilization numbers. Pillar 2 (detection quality) is partially independent: Table III reports TPR/PR-AUC against ground-truth attack labels for the 8-feature hardware prototype and the simulated 21-feature variant, and those results could have been poor if the leaf quantization or burst-flow mapping had distorted the teacher. However, the distillation-fidelity evidence (Eq. 5, Fig. 15, and the 'retains high TNR from Magnifier' claim) is self-referential: leaf labels are constructed from the autoencoder's reconstruction-error thresholds, so agreement with that autoencoder measures approximation fidelity, not independent predictive power. There is no load-bearing self-citation: borrowed components (bi-hash, double hash table, hypercube/whitelist-rule strategy) are cited to HorusEye [24] and other external works, not to the author's own prior results. No uniqueness theorem is imported. The FPR-budget mismatch between the 8-feature hardware variant (<=2e-3/<=1e-2) and the baselines (<=5e-5/<=5e-4) is a comparison-validity concern, not circularity. Overall: one localized self-definitional step in the consistency/fidelity argument, with the central deployment and ground-truth results retaining independent content, giving score 4.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

The central claim relies on a moderately large set of tunable hyperparameters and on several domain assumptions about traffic and switch hardware. The most under-validated component is the burst-to-flow mapping module, which is necessary to bridge burst-level and flow-level features during distillation.

free parameters (9)
  • Number of iTrees (t) = 200 (typically; 300 in some experiments)
    Controls how finely the feature space is divided; chosen based on consistency and detection experiments.
  • Sub-sample size (Psi) = 5000 (main), 400/256 in some experiments
    Controls iTree depth and feature boundary narrowness; tuned via consistency curves.
  • Data augmentation factor (k) = 50 (main), 150 in some experiments
    Number of uniformly sampled points per leaf used to compute expected reconstruction error; higher k improves distillation consistency up to a point.
  • Burst segmentation threshold (Nthreshold) = 15 (main), 4 for UNSW
    Maximum packets per burst; set based on distribution of burst lengths.
  • Idle timeout (delta_idle) = 1 second
    Inter-packet delay threshold defining a burst; taken from prior work.
  • Active timeout (delta_active) = 15 seconds
    Maximum burst duration; chosen to limit memory.
  • RMSE thresholds (T_u) = Set to meet FPR targets (e.g., 5e-5, 5e-4, 2e-3)
    Thresholds for each autoencoder used to label leaves; chosen per evaluation budget.
  • Autoencoder weights (w_u) = Equal weights (unspecified in text)
    Weights for combining autoencoder labels in Eq. 3; no detailed fitting is reported.
  • Profiler reward alpha = 0.5
    Trade-off between detection reward and memory footprint in Eq. 7.
assumptions (6)
  • domain assumption Normal (benign) traffic is available for training and is representative of the benign distribution at deployment time.
    The whole unsupervised method trains on normal traffic only; if the benign training data does not match deployment, detection degrades. Stated in threat model and dataset sections.
  • domain assumption Attacks from compromised IoT devices leave detectable traces in IP-layer features (packet sizes, timing, ports, TTL).
    The system uses only BL/PL features from IP layer; attacks like eavesdropping or MAC spoofing are excluded by threat model (§II-E).
  • standard math The iForest anomaly score threshold of 0.5 separates normal and anomalous samples.
    Standard iForest practice from Liu et al. [40]; used in Eq. 4 when combining original and distilled labels.
  • domain assumption The generated whitelist rules fit within the switch's memory and action constraints.
    The paper reports 9 stages and low TCAM/SRAM use (§VII-C), but assumes this generalizes to larger rule sets and other switches.
  • domain assumption The teacher autoencoders (Magnifier/Kitsune) are effective unsupervised anomaly detectors on the chosen features.
    The distilled iForest inherits the teacher's knowledge; if the teacher is weak, the student cannot exceed it. The paper uses state-of-the-art Magnifier.
  • ad hoc to paper The burst-to-flow mapping module accurately maps burst features to flow features.
    This module is introduced in supplementary §XIV, trained with mapping and reconstruction losses, and is not independently evaluated. The whole distillation from FL-based Magnifier to BL-based iForest depends on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CyberSentinel: Efficient Anomaly Detection in Programmable Switch using Knowledge Distillation." pith.science (2026). https://pith.science/paper/T2EMGPJD

@misc{pith2026241216693,
  author       = {Pith},
  title        = {Pith review of: CyberSentinel: Efficient Anomaly Detection in Programmable Switch using Knowledge Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2EMGPJD}},
  note         = {Machine review of arXiv:2412.16693}
}
abstract

The increasing volume of traffic (especially from IoT devices) is posing a challenge to the current anomaly detection systems. Existing systems are forced to take the support of the control plane for a more thorough and accurate detection of malicious traffic (anomalies). This introduces latency in making decisions regarding fast incoming traffic and therefore, existing systems are unable to scale to such growing rates of traffic. In this paper, we propose CyberSentinel, a high throughput and accurate anomaly detection system deployed entirely in the programmable switch data plane; making it the first work to accurately detect anomalies at line speed. To detect unseen network attacks, CyberSentinel uses a novel knowledge distillation scheme that incorporates "learned" knowledge of deep unsupervised ML models (\textit{e.g.}, autoencoders) to develop an iForest model that is then installed in the data plane in the form of whitelist rules. We implement a prototype of CyberSentinel on a testbed with an Intel Tofino switch and evaluate it on various real-world use cases. CyberSentinel yields similar detection performance compared to the state-of-the-art control plane solutions but with an increase in packet-processing throughput by $66.47\%$ on a $40$ Gbps link, and a reduction in average per-packet latency by $50\%$.

Figures

Figures reproduced from arXiv: 2412.16693 by the authors.

Figure 1
Figure 1. Protocol Independent Switch Architecture (PISA). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of CyberSentinel Data plane module. This module first extracts burst-level features from the incoming traffic/packets using the burst feature extractor (§IV-B). It detects anomalies based on the knowledge distilled iForest model (knowledge distillation from an ensemble of autoencoders) deployed in the form of whitelist rules. Details of knowledge distillation and whitelist rules generation are presented in … view at source ↗
Figure 3
Figure 3. Knowledge distillation of autoencoders into iForest and whitelist rules generation. (1) We first train an ensemble of autoencoders and (2) collect reconstruction errors by feeding each training sample in the trained ensemble. (3) We then train iForest model itself. We then (4) map each leaf node of trained iForest’s iTrees with the respective training samples. (5) Next, we embed each leaf node with expected reconstr… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Effect of various hyperparameters on consistency of knowledge [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: TPR and TNR comparison of iForest, Magnifier [24] and Magnifier [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: As shown, 1 we first match the incoming packet’s 5-tuple in the blacklist match action table. If there is a match, we can simply drop the packet or redirect it for further analysis [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 6
Figure 6. Figure 6: Data plane implementation of CyberSentinel at a glance Listing 1: Approach followed by [24] 1 action resubmit (bit<7> resub_port) { 2 ig_tm_md.ucast_egress_port = resub_port; 3 } 4 apply{ 5 //If packet is resubmitted one 6 if(ig_intr_md.ingress_port == RESUBMIT_PORT){ …
Figure 7
Figure 7. Figure 7: Data plane management V. CONTROL PLANE MODULE We discuss two components of control plane module. A. Profiler The profiler ( [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison of detection performance (TPR and TNR) of [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Effect of data augmentation factor k on CyberSentinel’s TPR and TNR. Number of iTrees = 200, subsample size = 400, and contamination ratio of 0.18 on both w / port and w/o port. Switch memory overheads. CyberSentinel takes up 9 stages of the pipeline with TCAM utilizat…
Figure 10
Figure 10. Figure 10: Comparison of CyberSentinel with Gulliver Tunnel, Magnifier, and HorusEye in terms of ROCAUC and PRAUC on UNSW NB-15 dataset. E. Detection performance on other datasets Setting. We use the popular intrusion detection dataset called UNSW NB-15 [45] from where we pick u…
Figure 12
Figure 12. Figure 12: Throughput and detection capability. G = Gulliver Tunnel, K = [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 14
Figure 14. Figure 14: We see a drop of 0.6% and 0.8% for TPR and TNR [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 13
Figure 13. Figure 13: P4 Code for average/variance/standard deviation of packet size [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Drop in TPR/TNR and increase in TCAM, SRAM, and ALUs resource [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: Effect of various hyperparameters on consistency of knowledge distillation algorithm. [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]
Figure 16
Figure 16. Figure 16: Effect of flow abnormality threshold on TPR and TNR of [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 17
Figure 17. Figure 17: Comparison of Magnifier-distilled iForest with iForest and Magnifier [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: Burst-to-flow mapping module (M) design For a given burst Bi having ni packets, the corresponding burst vector BVi is, BVi = ⟨ni , bf1, bf2, ..., bfd⟩ (13) where bfj is the j th burst-level (BL) feature and d is number of BL features. A sequence of burst vectors (for …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

82 extracted references · 69 canonical work pages

  1. [1]

    Gsma intelligence. the mobile economy 2020

    “Gsma intelligence. the mobile economy 2020.” [Online]. Available: https://www.gsmaintelligence.com,2020

  2. [2]

    Barefoot networks, tofino switch,

    “Barefoot networks, tofino switch,” 2021

  3. [3]

    Flowrest: Practical flow- level inference in programmable switches with random forests,

    A. T.-J. Akem, M. Gucciardo, and M. Fiore, “Flowrest: Practical flow- level inference in programmable switches with random forests,” IEEE INFOCOM 2023 - IEEE Conference on Computer Communications , 2023

  4. [4]

    Henna: Hierarchical machine learning inference in programmable switches,

    A. T.-J. Akem, B. B ¨ut¨un, M. Gucciardo, and M. Fiore, “Henna: Hierarchical machine learning inference in programmable switches,” in Proceedings of the 1st International Workshop on Native Network Intelligence, 2022

  5. [5]

    Jewel: Resource-efficient joint packet and flow level inference in pro- grammable switches,

    A. T.-J. Akem, B. B ¨ut¨un, M. Gucciardo, M. Fiore et al. , “Jewel: Resource-efficient joint packet and flow level inference in pro- grammable switches,” in IEEE International Conference on Computer Communications, 2024

  6. [6]

    Encrypted traffic classifi- cation at line rate in programmable switches with machine learning,

    A. T.-J. Akem, G. Fraysse, M. Fiore et al., “Encrypted traffic classifi- cation at line rate in programmable switches with machine learning,” in IEEE/IFIP Network Operations and Management Symposium , 2024

  7. [7]

    99% false positives: A qualitative study of SOC analysts’ perspectives on security alarms,

    B. A. Alahmadi, L. Axon, and I. Martinovic, “99% false positives: A qualitative study of SOC analysts’ perspectives on security alarms,” in 31st USENIX Security Symposium (USENIX Security 22) . Boston, MA: USENIX Association, Aug. 2022, pp. 2783–2800. [Online]. Available: https://www.usenix.org/conference/ usenixsecurity22/presentation/alahmadi

  8. [8]

    Aggregate- Based Congestion Control for Pulse-Wave DDoS Defense,

    A. G. Alcoz, M. Strohmeier, V . Lenders, and L. Vanbever, “Aggregate- Based Congestion Control for Pulse-Wave DDoS Defense,” in ACM SIGCOMM, Amsterdam, The Netherlands, August 2022. [Online]. Available: https://github.com/nsg-ethz/ACC-Turbo

Show all 82 references
  1. [9]

    Sok: Security evaluation of home-based iot deployments,

    O. Alrawi, C. Lever, M. Antonakakis, and F. Monrose, “Sok: Security evaluation of home-based iot deployments,” in 2019 IEEE symposium on security and privacy (sp) . IEEE, 2019, pp. 1362–1380

  2. [10]

    Insomnia: Towards concept-drift robustness in network intrusion detection,

    G. Andresini, F. Pendlebury, F. Pierazzi, C. Loglisci, A. Appice, and L. Cavallaro, “Insomnia: Towards concept-drift robustness in network intrusion detection,” in Proceedings of the 14th ACM workshop on artificial intelligence and security , 2021, pp. 111–122

  3. [11]

    Sok: The impact of unlabelled data in cyberthreat detection,

    G. Apruzzese, P. Laskov, and A. Tastemirova, “Sok: The impact of unlabelled data in cyberthreat detection,” in 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P) . IEEE, 2022, pp. 20–42

  4. [12]

    Dos and don’ts of machine learning in computer security,

    D. Arp, E. Quiring, F. Pendlebury, A. Warnecke, F. Pierazzi, C. Wress- negger, L. Cavallaro, and K. Rieck, “Dos and don’ts of machine learning in computer security,” in 31st USENIX Security Symposium (USENIX Security 22), 2022, pp. 3971–3988

  5. [13]

    Rectified decision trees: Towards interpretability, compression and empirical soundness,

    J. Bai, Y . Li, J. Li, Y . Jiang, and S. Xia, “Rectified decision trees: Towards interpretability, compression and empirical soundness,” arXiv preprint arXiv:1903.05965, 2019

  6. [14]

    Multinomial random forest,

    J. Bai, Y . Li, J. Li, X. Yang, Y . Jiang, and S.-T. Xia, “Multinomial random forest,” Pattern Recognition, vol. 122, p. 108331, 2022

  7. [15]

    Flowlens: Enabling efficient flow classification for ml- based network security applications

    D. Barradas, N. Santos, L. Rodrigues, S. Signorello, F. M. Ramos, and A. Madeira, “Flowlens: Enabling efficient flow classification for ml- based network security applications.” in NDSS, 2021

  8. [16]

    Providing iot host-based datasets for intrusion detection research,

    V . H. Bezerra, V . G. T. da Costa, R. A. Martins, S. B. Junior, R. S. Miani, and B. B. Zarpelao, “Providing iot host-based datasets for intrusion detection research,” in Anais do XVIII Simp ´osio Brasileiro de Seguranc ¸a da Informac ¸ao e de Sistemas Computacionais, 2018

  9. [17]

    P4: Pro- gramming protocol-independent packet processors,

    P. Bosshart, D. Daly, G. Gibb, M. Izzard, N. McKeown, J. Rexford, C. Schlesinger, D. Talayco, A. Vahdat, G. Varghese et al. , “P4: Pro- gramming protocol-independent packet processors,” ACM SIGCOMM Computer Communication Review , 2014

  10. [18]

    On the evaluation of unsupervised outlier detection: measures, datasets, and an empirical study,

    G. O. Campos, A. Zimek, J. Sander, R. J. Campello, B. Micenkov ´a, E. Schubert, I. Assent, and M. E. Houle, “On the evaluation of unsupervised outlier detection: measures, datasets, and an empirical study,” Data mining and knowledge discovery , vol. 30, pp. 891–927, 2016

  11. [19]

    P4 architectures,

    C. Cascaval and D. Daly, “P4 architectures,” 2021

  12. [20]

    Iotguard: Dynamic enforce- ment of security and safety policy in commodity iot

    Z. B. Celik, G. Tan, and P. D. McDaniel, “Iotguard: Dynamic enforce- ment of security and safety policy in commodity iot.” in NDSS, 2019

  13. [21]

    drmt: Disaggregated programmable switching,

    S. Chole, A. Fingerhut, S. Ma, A. Sivaraman, S. Vargaftik, A. Berger, G. Mendelson, M. Alizadeh, S.-T. Chuang, I. Keslassy et al. , “drmt: Disaggregated programmable switching,” in Proceedings of the Con- ference of the ACM Special Interest Group on Data Communication , 2017

  14. [22]

    Cobzas ¸, R

    S ¸. Cobzas ¸, R. Miculescu, A. Nicolae et al. , Lipschitz functions . Springer, 2019, vol. 2019936365

  15. [23]

    Iot malware,

    F. Ding, “Iot malware,” 2017. [Online]. Available: https://github.com/ ifding/iot-malware

  16. [24]

    HorusEye: A realtime IoT malicious traffic detection framework using programmable switches,

    Y . Dong, Q. Li, K. Wu, R. Li, D. Zhao, G. Tyson, J. Peng, Y . Jiang, S. Xia, and M. Xu, “HorusEye: A realtime IoT malicious traffic detection framework using programmable switches,” in 32nd USENIX Security Symposium (USENIX Security 23) . Anaheim, CA: USENIX Association, Aug....

  17. [25]

    Building a better netflow,

    C. Estan, K. Keys, D. Moore, and G. Varghese, “Building a better netflow,” ACM SIGCOMM Computer Communication Review , vol. 34, no. 4, pp. 245–256, 2004

  18. [26]

    A learning methodology for line-rate ransomware mitigation with p4 switches,

    K. Friday, E. Bou-Harb, and J. Crichigno, “A learning methodology for line-rate ransomware mitigation with p4 switches,” in International Conference on Network and System Security , 2022

  19. [27]

    Inc: In-network classification of botnet propagation at line rate,

    K. Friday, E. Kfoury, E. Bou-Harb, and J. Crichigno, “Inc: In-network classification of botnet propagation at line rate,” in European Sympo- sium on Research in Computer Security , 2022

  20. [28]

    Distilling a neural network into a soft decision tree,

    N. Frosst and G. Hinton, “Distilling a neural network into a soft decision tree,” arXiv preprint arXiv:1711.09784 , 2017

  21. [29]

    Realtime robust malicious traffic detection via frequency domain analysis,

    C. Fu, Q. Li, M. Shen, and K. Xu, “Realtime robust malicious traffic detection via frequency domain analysis,” in Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security , 2021, pp. 3431–3446

  22. [30]

    Nsa- net: A netflow sequence attention network for virtual private network traffic detection,

    P. Fu, C. Liu, Q. Yang, Z. Li, G. Gou, G. Xiong, and Z. Li, “Nsa- net: A netflow sequence attention network for virtual private network traffic detection,” in Web Information Systems Engineering–WISE 2020: 16 21st International Conference, Amsterdam, The Netherlands, October 2...

  23. [31]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville, Deep learning. MIT press, 2016

  24. [32]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” International Journal of Computer Vision , vol. 129, no. 6, pp. 1789–1819, 2021

  25. [33]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

  26. [34]

    Leo: Online ml-based traffic classification at multi-terabit line rate

    S. U. Jafri, S. Rao, V . Shrivastav, and M. Tawarmalani, “Leo: Online ml-based traffic classification at multi-terabit line rate.” NSDI, 2024

  27. [35]

    Transcend: Detecting concept drift in malware classification models,

    R. Jordaney, K. Sharad, S. K. Dash, Z. Wang, D. Papini, I. Nouretdinov, and L. Cavallaro, “Transcend: Detecting concept drift in malware classification models,” in 26th USENIX security symposium (USENIX security 17), 2017, pp. 625–642

  28. [36]

    Lipschitz continuous autoencoders in application to anomaly detection,

    Y .-g. Kim, Y . Kwon, H. Chang, and M. C. Paik, “Lipschitz continuous autoencoders in application to anomaly detection,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2020, pp. 2507–2517

  29. [37]

    Towards the development of realistic botnet dataset in the internet of things for network forensic analytics: Bot-iot dataset,

    N. Koroniotis, N. Moustafa, E. Sitnikova, and B. Turnbull, “Towards the development of realistic botnet dataset in the internet of things for network forensic analytics: Bot-iot dataset,” Future Generation Computer Systems, 2019

  30. [38]

    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 , 2020

  31. [39]

    Fs-net: A flow sequence network for encrypted traffic classification,

    C. Liu, L. He, G. Xiong, Z. Cao, and Z. Li, “Fs-net: A flow sequence network for encrypted traffic classification,” in IEEE INFOCOM 2019- IEEE Conference On Computer Communications . IEEE, 2019, pp. 1171–1179

  32. [40]

    Isolation forest,

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in 2008 eighth ieee international conference on data mining . IEEE, 2008, pp. 413– 422

  33. [41]

    Jaqen: A high-performance switch-native approach for detecting and mitigating volumetric ddos attacks with programmable switches

    Z. Liu, H. Namkung, G. Nikolaidis, J. Lee, C. Kim, X. Jin, V . Braver- man, M. Yu, and V . Sekar, “Jaqen: A high-performance switch-native approach for detecting and mitigating volumetric ddos attacks with programmable switches.” in USENIX Security Symposium , 2021, pp. 3829–3846

  34. [42]

    Pinpointing hidden iot devices via spatial-temporal traffic fingerprinting,

    X. Ma, J. Qu, J. Li, J. C. Lui, Z. Li, and X. Guan, “Pinpointing hidden iot devices via spatial-temporal traffic fingerprinting,” in IEEE INFO- COM 2020-IEEE Conference on Computer Communications , 2020

  35. [43]

    Silkroad: Making stateful layer-4 load balancing fast and cheap using switching asics,

    R. Miao, H. Zeng, C. Kim, J. Lee, and M. Yu, “Silkroad: Making stateful layer-4 load balancing fast and cheap using switching asics,” in Proceedings of the Conference of the ACM Special Interest Group on Data Communication, 2017

  36. [44]

    Kitsune: an ensemble of autoencoders for online network intrusion detection,

    Y . Mirsky, T. Doitshman, Y . Elovici, and A. Shabtai, “Kitsune: an ensemble of autoencoders for online network intrusion detection,” Network and Distributed System Security Symposium 2018 (NDSS’18) , 2018

  37. [45]

    Unsw-nb15: a comprehensive data set for network intrusion detection systems (unsw-nb15 network data set),

    N. Moustafa and J. Slay, “Unsw-nb15: a comprehensive data set for network intrusion detection systems (unsw-nb15 network data set),” in 2015 military communications and information systems conference (MilCIS). IEEE, 2015, pp. 1–6

  38. [46]

    Persistent ospf attacks

    G. Nakibly, A. Kirshon, D. Gonikman, and D. Boneh, “Persistent ospf attacks.” in NDSS, 2012

  39. [47]

    Polygraph: Automatically gener- ating signatures for polymorphic worms,

    J. Newsome, B. Karp, and D. Song, “Polygraph: Automatically gener- ating signatures for polymorphic worms,” in 2005 IEEE Symposium on Security and Privacy (S&P’05) . IEEE, 2005, pp. 226–241

  40. [48]

    Sailfish: Accelerating cloud-scale multi-tenant multi- service gateways with programmable switches,

    T. Pan, N. Yu, C. Jia, J. Pi, L. Xu, Y . Qiao, Z. Li, K. Liu, J. Lu, J. Lu et al. , “Sailfish: Accelerating cloud-scale multi-tenant multi- service gateways with programmable switches,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference , 2021, pp. 194–206

  41. [49]

    Learning representations of ultrahigh-dimensional data for random distance-based outlier detection,

    G. Pang, L. Cao, L. Chen, and H. Liu, “Learning representations of ultrahigh-dimensional data for random distance-based outlier detection,” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , 2018, pp. 2041–2050

  42. [50]

    An {Input-Agnostic} hierarchical deep learning frame- work for traffic fingerprinting,

    J. Qu, X. Ma, J. Li, X. Luo, L. Xue, J. Zhang, Z. Li, L. Feng, and X. Guan, “An {Input-Agnostic} hierarchical deep learning frame- work for traffic fingerprinting,” in 32nd USENIX Security Symposium (USENIX Security 23) , 2023, pp. 589–606

  43. [51]

    Firma: Malware clustering and network signature generation with mixed network behaviors,

    M. Z. Rafique and J. Caballero, “Firma: Malware clustering and network signature generation with mixed network behaviors,” in Research in Attacks, Intrusions, and Defenses: 16th International Symposium, RAID 2013, Rodney Bay, St. Lucia, October 23-25, 2013. Proceedings 16 . Sp...

  44. [52]

    Less is more: Building selective anomaly ensembles,

    S. Rayana and L. Akoglu, “Less is more: Building selective anomaly ensembles,” Acm transactions on knowledge discovery from data (tkdd), vol. 10, no. 4, pp. 1–33, 2016

  45. [53]

    Marina: Realizing ml-driven real-time network traffic monitoring at terabit scale,

    M. Seufert, K. Dietz, N. Wehner, S. Geißler, J. Sch ¨uler, M. Wolz, A. Hotho, P. Casas, T. Hoßfeld, and A. Feldmann, “Marina: Realizing ml-driven real-time network traffic monitoring at terabit scale,” IEEE Transactions on Network and Service Management , 2024

  46. [54]

    Evaluating the power of flexible packet processing for network resource allocation,

    N. K. Sharma, A. Kaufmann, T. Anderson, A. Krishnamurthy, J. Nelson, and S. Peter, “Evaluating the power of flexible packet processing for network resource allocation,” in14th USENIX Symposium on Networked Systems Design and Implementation (NSDI 17) , 2017

  47. [55]

    Classifying iot devices in smart environments using network traffic characteristics,

    A. Sivanathan, H. H. Gharakheili, F. Loi, A. Radford, C. Wijenayake, A. Vishwanath, and V . Sivaraman, “Classifying iot devices in smart environments using network traffic characteristics,” IEEE Transactions on Mobile Computing , 2018

  48. [56]

    Packet transactions: High-level programming for line-rate switches,

    A. Sivaraman, A. Cheung, M. Budiu, C. Kim, M. Alizadeh, H. Balakr- ishnan, G. Varghese, N. McKeown, and S. Licking, “Packet transactions: High-level programming for line-rate switches,” in Proceedings of the 2016 ACM SIGCOMM Conference , 2016

  49. [57]

    Heavy-hitter detection entirely in the data plane,

    V . Sivaraman, S. Narayana, O. Rottenstreich, S. Muthukrishnan, and J. Rexford, “Heavy-hitter detection entirely in the data plane,” in Proceedings of the Symposium on SDN Research , 2017, pp. 164–176

  50. [58]

    Scaling hardware accelerated network monitoring to concurrent and dynamic queries with *flow,

    J. Sonchack, O. Michel, A. J. Aviv, E. Keller, and J. M. Smith, “Scaling hardware accelerated network monitoring to concurrent and dynamic queries with *flow,” in 2018 USENIX Annual Technical Conference (USENIX ATC 18) , 2018, pp. 823–835

  51. [59]

    Improved semantic represen- tations from tree-structured long short-term memory networks,

    K. S. Tai, R. Socher, and C. D. Manning, “Improved semantic represen- tations from tree-structured long short-term memory networks,” ACL, 2015

  52. [60]

    Zerowall: Detecting zero-day web attacks through encoder-decoder recurrent neural networks,

    R. Tang, Z. Yang, Z. Li, W. Meng, H. Wang, Q. Li, Y . Sun, D. Pei, T. Wei, Y . Xuet al., “Zerowall: Detecting zero-day web attacks through encoder-decoder recurrent neural networks,” in IEEE INFOCOM 2020- IEEE Conference on Computer Communications , 2020

  53. [61]

    A lipschitz-constrained anomaly discriminator framework,

    A. Tong, G. Wolf, and S. Krishnaswamy, “A lipschitz-constrained anomaly discriminator framework,” arXiv preprint arXiv:1905.10710 , 2019

  54. [62]

    Fixing bias in reconstruction-based anomaly detection with lipschitz discriminators,

    ——, “Fixing bias in reconstruction-based anomaly detection with lipschitz discriminators,” Journal of Signal Processing Systems, vol. 94, no. 2, pp. 229–243, 2022

  55. [63]

    Packet-level signatures for smart home devices,

    R. Trimananda, J. Varmarken, A. Markopoulou, and B. Demsky, “Packet-level signatures for smart home devices,” in Network and Distributed Systems Security (NDSS) Symposium , vol. 2020, 2020

  56. [64]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

  57. [65]

    Machine learning for networking: Workflow, advances and opportunities,

    M. Wang, Y . Cui, X. Wang, S. Xiao, and J. Jiang, “Machine learning for networking: Workflow, advances and opportunities,” Ieee Network, vol. 32, no. 2, pp. 92–99, 2017

  58. [66]

    Integer quantization for deep learning inference: Principles and empirical evaluation,

    H. Wu, P. Judd, X. Zhang, M. Isaev, and P. Micikevicius, “Integer quantization for deep learning inference: Principles and empirical evaluation,” arXiv preprint arXiv:2004.09602 , 2020

  59. [67]

    A short-term load forecasting method based on gru-cnn hybrid neural network model,

    L. Wu, C. Kong, X. Hao, and W. Chen, “A short-term load forecasting method based on gru-cnn hybrid neural network model,” Mathematical problems in engineering , vol. 2020, pp. 1–10, 2020. 17

  60. [68]

    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, 2022

  61. [69]

    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, 2024

  62. [70]

    {NetWarden}: Mitigating network covert channels while preserving performance,

    J. Xing, Q. Kang, and A. Chen, “ {NetWarden}: Mitigating network covert channels while preserving performance,” in 29th USENIX Secu- rity Symposium (USENIX Security 20) , 2020, pp. 2039–2056

  63. [71]

    Anomaly network traffic detection based on deep transfer learning,

    P. Xiong, B. Cui, and Z. Cheng, “Anomaly network traffic detection based on deep transfer learning,” in Innovative Mobile and Internet Ser- vices in Ubiquitous Computing: Proceedings of the 14th International Conference on Innovative Mobile and Internet Services in Ubiquitous ...

  64. [72]

    Poseidon: Mitigating volumetric ddos attacks with programmable switches,

    M. Zhang, G. Li, S. Wang, C. Liu, A. Chen, H. Hu, G. Gu, Q. Li, M. Xu, and J. Wu, “Poseidon: Mitigating volumetric ddos attacks with programmable switches,” in the 27th Network and Distributed System Security Symposium (NDSS 2020) , 2020

  65. [73]

    Homonit: Monitoring smart home apps from encrypted traffic,

    W. Zhang, Y . Meng, Y . Liu, X. Zhang, Y . Zhang, and H. Zhu, “Homonit: Monitoring smart home apps from encrypted traffic,” in Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security, 2018

  66. [74]

    pheavy: Predicting heavy flows in the programmable data plane,

    X. Zhang, L. Cui, F. P. Tso, and W. Jia, “pheavy: Predicting heavy flows in the programmable data plane,” IEEE Transactions on Network and Service Management , 2021

  67. [75]

    Iisy: Practical in-network classification,

    C. Zheng, Z. Xiong, T. T. Bui, S. Kaupmees, R. Bensoussane, A. Bern- abeu, S. Vargaftik, Y . Ben-Itzhak, and N. Zilberman, “Iisy: Practical in-network classification,” arXiv preprint arXiv:2205.08243 , 2022

  68. [76]

    Automating in-network machine learning,

    C. Zheng, M. Zang, X. Hong, R. Bensoussane, S. Vargaftik, Y . Ben- Itzhak, and N. Zilberman, “Automating in-network machine learning,” arXiv preprint arXiv:2205.08824 , 2022

  69. [77]

    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

  70. [78]

    An efficient design of intelligent network data plane,

    G. Zhou, Z. Liu, C. Fu, Q. Li, and K. Xu, “An efficient design of intelligent network data plane,” in 32nd USENIX Security Symposium (USENIX Security 23). Anaheim, CA: USENIX Association , 2023

  71. [79]

    Cerberus: Enabling efficient and effective in- network monitoring on programmable switches,

    H. Zhou and G. Gu, “Cerberus: Enabling efficient and effective in- network monitoring on programmable switches,” in 2024 IEEE Sympo- sium on Security and Privacy (SP) , 2023. 18 SUPPLEMENTARY MATERIAL X. D ISTILLED I FOREST MODEL AND RULES GENERATION We present a detailed proc...

  72. [80]

    ( Label Consistency ). Proof. It is trivial to see that for each iTree, labels of all samples in an iTree hypercube are consistent. Following the merging of iTree hypercubes into iForest hypercubes, all sample points inside an iForest hypercube must belong to the same iTree hy...

  73. [81]

    For this experiment, we do not consider anomaly score from the original iForest

    into the iForest. For this experiment, we do not consider anomaly score from the original iForest. In other words, the label of distilled iForest can be obtained for this experiment by simply traversing all t iTrees for a sample x and taking a majority vote of the labels obtai...

  74. [150]

    The setting for the Magnifier is the same as in [24]. Metrics. We use two metrics: PR AUC and ROCAUC. The ROC curve indicates the true positives against false positives, while the PR curve summarises precision and recall of the anomaly class only. Implementation. We only consi...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.