Pith. sign in

REVIEW 3 major objections 4 minor 68 references

Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that Pegasus is the first intelligent-dataplane design that runs multiple deep-learning models—MLP, RNN, CNN, and autoencoders—on commodity programmable switches, with an average 97.3% classification accuracy on three…

desk verdict A real Tofino 2 DL inference framework worth reviewing, but its headline accuracy numbers rest on a changed model architecture and uncovered baselines. read the letter →

arxiv 2506.05779 v1 pith:F722GZCV submitted 2025-06-06 cs.NI cs.LG

Yinchao Zhang , Su Yao , Yong Feng , Kang Chen , Tong Li , Zhuotao Liu , Yi Zhao , Lexuan Zhang
show 4 more authors
Xiangyu Gao Feng Xiong Qi Li Ke Xu
This is my paper
classification cs.NIcs.LG
keywords intelligentdataplaneprogrammableswitchesP4match-actiontablesdeeplearninginferencefuzzymatchingtrafficclassificationin-networkmachineautoencoderanomalydetection
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

Pegasus claims that deep learning inference can run on commodity programmable switches at line rate if network operators stop trying to compute DL operations and instead translate them into three primitives: Partition, Map, and SumReduce. The load-bearing idea is fuzzy matching, which replaces arithmetic with table lookups over cluster centroids so that a P4 match-action pipeline can evaluate matrix products, convolutions, activations, and normalizations without multiplication or floating point. On three traffic datasets the design supports 3840-bit inputs and a 6083 Kb model, achieves an average macro-F1 of 97.3%, and stays within 0.2–1.7% of full-precision CPU/GPU inference while improving throughput by more than 3800× over CPU. The paper argues this makes line-rate DL practical on existing switch hardware, which would matter for in-network traffic classification, anomaly detection, and other terabit-scale intelligent services.

What carries the argument

The machinery is a three-primitive decomposition of DL operators—Partition, Map, and SumReduce—compiled onto the match-action table abstraction. The load-bearing component is fuzzy matching: a clustering tree whose internal nodes are feature-threshold comparisons maps an input sub-vector to a leaf centroid index, and a single table lookup returns the approximate output of the target function on that sub-vector. Greedy minimum-SSE splitting builds the tree, backpropagation through the differentiable representation of the routing step fine-tunes centroids and thresholds, and primitive fusion (linear reordering, merging consecutive Maps, and the Neural-Additive-Model-style removal of intermediate SumReduces) compresses whole pipelines into as few as one or two lookups.

What would settle it

Take a trained CNN deployed through Pegasus, collect test inputs that lie exactly on the feature-threshold boundaries of the fuzzy-matching tree, and compare the switch's macro-F1 with full-precision CPU inference; if the boundary-adjacent subset shows accuracy loss much larger than the reported average 1.08%, the continuity premise is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the match-action table abstraction—previously used only for exact packet matching—can be made to carry deep learning inference by decomposing every DL operator into Partition, Map, and SumReduce and serving the Map step with fuzzy-matching lookup tables. Each input sub-vector is routed through a feature-threshold clustering tree to a leaf centroid; the centroid's precomputed, full-precision result is fetched in one lookup, and SumReduce aggregates the partial results. Because table size grows with the number of clusters rather than the product of input bit widths, the design can afford full-precision weights and 8-bit fixed-point activations, and backpropagation fine-tunes the centroids and tree parameters to keep approximation error small. The reported result is a dataplane that classifies raw packet sequences as well as full-precision models do, with CNN-L reaching 99.66%, 93.80%, and 98.72% macro-F1 on PeerRush, CICIOT, and ISCXVPN respectively.

Load-bearing premise

The accuracy claims rest on the assumption that DL operators are smooth enough that replacing an input sub-vector with its cluster centroid leaves the final output almost unchanged; if that fails, backprop fine-tuning can only repair the smooth part of the error.

Editorial extensions

If this is right

  • If Pegasus is correct, a commodity PISA/Tofino-2 switch can run supervised DL classifiers on raw packet sequences at line rate, not just tree-based models.
  • The reported accuracy gap of 0.2–1.7% versus full-precision CPU/GPU inference means traffic-analysis tasks previously run off-switch could move onto the dataplane without a noticeable accuracy hit.
  • The 3840-bit input and 6083 Kb model numbers imply dataplane models can grow well past the 18–128-bit and roughly 25 Kb limits of earlier in-network approaches.
  • The autoencoder evaluation suggests unsupervised anomaly detection, including zero-day and DoS traffic, can be performed on the switch with per-flow state as small as 28 bits.
  • Throughput gains of over 3800× versus CPU and 600× versus GPU, even in an idealized comparison, indicate that the main benefit of dataplane DL is not accuracy but sustained line-rate processing.

Reading between the lines

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

  • Editorial extension: fuzzy matching is a piecewise-constant approximation of each operator, so accuracy should degrade most on inputs that sit near cluster boundaries or far from the training distribution; a boundary-focused stress test would be a sharper test than average F1.
  • Editorial extension: because the cost of a lookup grows with cluster count rather than operand bit width, the same decomposition could be pushed toward larger models or higher-dimensional inputs if switch SRAM/TCAM budgets grow, but the practical ceiling is set by memory bandwidth and the continuity assumption.
  • Editorial extension: the Partition–Map–SumReduce decomposition is general enough to apply to other in-network machine learning domains, such as routing, congestion control, or telemetry, wherever the operator can be approximated by lookup tables.
  • Editorial extension: the paper explicitly notes transformers cannot fit in current resources; an obvious extension is to test whether attention layers can be approximated by the same primitives when fused at block level.
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. The paper presents Pegasus, a framework for deep learning inference on programmable dataplane switches. DL operators are decomposed into three primitives—Partition, Map, and SumReduce—where Map uses fuzzy matching with clustering-tree-based lookup tables. Primitive Fusion (basic and advanced) and adaptive fixed-point quantization are introduced to reduce table lookups and improve numerical accuracy. The authors implement Pegasus on a Barefoot Tofino 2 switch and report evaluations for MLP, RNN, CNN, and AutoEncoder models on three traffic classification datasets, claiming up to 97.3% average accuracy, 3840-bit input scale, 6083 Kb model size, and up to 22.8% accuracy improvement over prior systems.

Significance. If the claims hold, Pegasus is a noteworthy systems contribution: it demonstrates that multiple DL model families can be mapped onto commodity P4 switch primitives with modest accuracy loss and gives concrete resource-utilization measurements on Tofino 2. The paper ships a real hardware implementation, reports accuracy on held-out test splits with standard supervised training, and provides falsifiable resource tables. The main scientific value is the demonstration that MAT-based fuzzy matching plus SumReduce can approximate a range of DL operators at line rate. However, the interpretation of the central CNN results is currently ambiguous because the reported CNN models have undergone architectural modification before deployment, and the headline comparisons against prior work are not controlled for input scale and model capacity.

major comments (3)
  1. [§6.3, §7.5, Table 5, Figure 9] The reported CNN-L accuracy is not interpretable as preserving CNN inference unless the CPU/GPU baseline uses the same architecture as the model actually running on the dataplane. Section 4.3 defines Advanced Primitive Fusion as either removing all nonlinear mappings or reducing the model to a single final SumReduce (a Neural Additive Model), and Section 6.3 states that CNN-M, CNN-L, and AutoEncoder applied this technique. If the dataplane CNN-L is a NAM-like additive model over 8-packet windows, then the F1 scores of 99.66/93.80/98.72 in Table 5 and the 0.2-0.9% gap to GPU/CPU in Figure 9 do not demonstrate that Pegasus preserves CNN inference at 3840-bit scale; they demonstrate that an additive model over the same input works on these datasets. The paper must disclose whether the GPU/CPU baseline is the original textcnn or the post-fusion architecture, and ideally report all three variants: original textcnn on CPU/GPU, post-fusion full-precision model on CPU/GPU, and Pegasus on the dataplane. Without this, the central claim conflates fuzzy-matching error with a fundamental change of model.
  2. [§7.3, Table 5, Table 2] The headline comparisons 'average accuracy gains of 17.2%, 22.8%, and 17.9% over Leo, N3IC, and BoS' are not controlled. CNN-L uses a 3840-bit input and 6083 Kb model, whereas N3IC uses a 128-bit input and 24.4 Kb model, and BoS uses an 18-bit input and 25.6 Kb model. The accuracy improvement may therefore be due to larger input scale and model capacity rather than to Pegasus's design. Please report results for Pegasus variants whose input scale and model capacity are matched to each baseline, and state explicitly which comparisons are controlled. The presentation in Table 2 of '22.8%↑' without noting the scale difference is misleading.
  3. [§7.3, Figure 7, Table 6] The per-flow storage accounting for CNN-L is inconsistent. The text says CNN-L uses 48 bits per flow, comprising 16 bits for the previous packet timestamp plus 4 bits for a fuzzy index per packet for 7 packets, which sums to 44 bits, not 48; Table 6 reports 44 bits. The footnote stating that four 4-bit registers were replaced by four 8-bit registers is also unclear. Please correct the numbers and clarify the register accounting, since the flow-scalability claim in Section 7.3 depends on this figure.
minor comments (4)
  1. [§6.3, Table 5] Section 6.3 refers to 'CNN-S extends CNN-B' and 'CNN-L builds on CNN-S', but Table 5 lists CNN-B, CNN-M, and CNN-L without any CNN-S. The naming should be made consistent throughout.
  2. [Abstract, §7.3] The abstract reports 'average accuracy improvement of up to 22.8%' and 'up to 248x larger model size and 212x larger input scale' without specifying that these figures refer to different baselines with very different input scales; this should be stated more precisely.
  3. [§4.2] The continuity assumption underlying fuzzy matching is asserted but not quantitatively validated; reporting the distribution of distances between inputs and their assigned centroids on the test sets would strengthen the argument.
  4. [§7.5, Figure 9] The throughput comparison in Figure 9(d) reports only idealized peak throughput with pre-loaded features; the text acknowledges this, but a brief statement of the actual end-to-end measurement conditions should appear in the caption.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: fuzzy-matching tables are trained on the training split and tested on held-out data; self-citations are baselines or subroutines, and the Advanced Fusion issue is a benchmarking confound, not circularity.

full rationale

Pegasus's derivation chain is self-contained as an empirical systems paper. The cluster-tree parameters and centroids that define Map lookups are learned from the training portion of each dataset and evaluated on a separate held-out test split (Section 7.1), which is the standard supervised-learning separation rather than a fitted input renamed as a prediction; the backpropagation fine-tuning in Section 4.4 likewise operates on the training data only. The claimed accuracy numbers in Table 5 are direct measurements on test flows, not quantities implied by the construction of the mapping tables. The only self-referential elements are citations to the authors' prior BoS and NetBeacon work: BoS [46] is used as a comparison baseline and as the starting point for the RNN-B model, and NetBeacon [58] supplies the CRC range-coding subroutine used to implement fuzzy-index lookups. Neither citation is load-bearing for the central claim that Pegasus is the first IDP design to support multiple DL model types; the claim rests on the P4 implementation and Table 5/Figure 9 measurements. The Advanced Primitive Fusion modification described in Section 4.3 does change CNN-M/CNN-L to a NAM-like additive form, and Section 7.5 does not state whether the GPU/CPU baseline is the original or post-fusion architecture; this is a legitimate external-validity/correctness concern about benchmark interpretation, but it is not a circular reduction because the reported F1 scores are empirical measurements of the deployed fused model, not derivations from the training fit.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The framework's accuracy depends on a set of trained parameters (cluster thresholds, centroids, fused table entries, fixed-point positions) and on the assumptions that DL operators are locally continuous, that the greedy clustering is near-optimal, and that PISA line-rate holds. No new physical entities are introduced; Partition, Map, and SumReduce are software abstractions with no independent empirical evidence beyond the paper's own evaluation.

free parameters (5)
  • Cluster tree thresholds and split features = Learned from training set via greedy min-SSE splits
    Section 4.2: each node selects optimal feature and threshold to minimize SSE; these determine the fuzzy index and hence which centroid approximates each input. Central to accuracy.
  • Cluster centroids = Mean vectors of clusters, fine-tuned by backpropagation
    Section 4.2 and 4.4: centroids are the stored values used by Map lookups; backprop adjusts them to reduce output error. This is the core approximation.
  • Fixed-point quantization positions = Pre-calculated per layer
    Section 4.4 Adaptive Fixed-Point Quantization: input and output fixed-point positions chosen to maximize bit-width utilization; affects numerical precision of SumReduce and activations.
  • Mapping table outputs (fused operator results) = Precomputed results of fused operations at centroids, refined by backprop
    Section 4.3 and 4.4: table entries encode f(centroid) for fused primitive sequences; after fine-tuning they are fitted values, not derived from first principles.
  • Partition size, clustering depth, CNN kernel and stride = Chosen per model; computed by translator
    PegasusSyntax parameters (Section 6.2) that determine the number and size of lookups; not systematically swept in the paper, so their sensitivity is unknown.
assumptions (6)
  • domain assumption DL operators are locally continuous: small input changes cause small output changes, so replacing an input with its cluster centroid is accurate.
    Section 4.2: 'This approach leverages the continuity of DL operators ... allowing minor variations in the input without significantly affecting the output [61]'. This is the load-bearing premise for fuzzy matching.
  • domain assumption Training data are i.i.d. and representative of test and deployment traffic, and the clustering tree built on training data generalizes online.
    Section 4.2: 'Based on the independent and identically distributed (i.i.d.) assumption of DL [40,41]'.
  • domain assumption Any P4 program compiled for a PISA pipeline runs at fixed line rate regardless of table complexity.
    Section 7.5: 'Since the PISA pipeline on programmable switches ensures that any program compiled for it can run at line-rate' supports the 3800x and 600x throughput claim.
  • standard math Backpropagation through the centroid assignment can be implemented via the matrix representation of decision trees from Zhang [51].
    Section 4.4 cites Zhang [51] for simulating centroid assignment through matrix operations; the paper does not derive or verify this representation.
  • domain assumption The greedy min-SSE clustering tree provides a sufficiently good quantization (near-optimal) for the dataplane approximation.
    Section 4.2: 'Although the greedy strategy does not guarantee a global optimum, it provides a near-optimal split'. The accuracy results depend on this.
  • standard math Consecutive numerical ranges can be converted to ternary TCAM rules via CRC without prohibitive expansion.
    Section 6.1 uses the Consecutive Range Coding algorithm [58] to implement the clustering tree comparator with range matching; the paper relies on this external technique.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane." pith.science (2026). https://pith.science/paper/F722GZCV

@misc{pith2026250605779,
  author       = {Pith},
  title        = {Pith review of: Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F722GZCV}},
  note         = {Machine review of arXiv:2506.05779}
}
read the original abstract

The paradigm of Intelligent DataPlane (IDP) embeds deep learning (DL) models on the network dataplane to enable intelligent traffic analysis at line-speed. However, the current use of the match-action table (MAT) abstraction on the dataplane is misaligned with DL inference, leading to several key limitations, including accuracy degradation, limited scale, and lack of generality. This paper proposes Pegasus to address these limitations. Pegasus translates DL operations into three dataplane-oriented primitives to achieve generality: Partition, Map, and SumReduce. Specifically, Partition "divides" high-dimensional features into multiple low-dimensional vectors, making them more suitable for the dataplane; Map "conquers" computations on the low-dimensional vectors in parallel with the technique of fuzzy matching, while SumReduce "combines" the computation results. Additionally, Pegasus employs Primitive Fusion to merge computations, improving scalability. Finally, Pegasus adopts full precision weights with fixed-point activations to improve accuracy. Our implementation on a P4 switch demonstrates that Pegasus can effectively support various types of DL models, including Multi-Layer Perceptron (MLP), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), and AutoEncoder models on the dataplane. Meanwhile, Pegasus outperforms state-of-the-art approaches with an average accuracy improvement of up to 22.8%, along with up to 248x larger model size and 212x larger input scale.

Figures

Figures reproduced from arXiv: 2506.05779 by the authors.

Figure 2
Figure 2. Implementation of a Map primitive: how in￾put sub-vector (3, 7) retrieves results (1.8, 4) as the ap￾proximation of 𝑓 (𝑋𝑖) = 0.4𝑋𝑖 + 1. 4 PEGASUS PRIMITIVES 4.1 Primitives Pegasus primitives fall into three categories: Partition, Map, and SumReduce, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Correspondence between the MAT abstraction [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Primitive Fusion techniques: ❶ Basic Prim￾itive Fusion; ❷ Advanced Primitive Fusion with Re￾moval of Nonlinear Mappings; ❸ Advanced Primitive Fusion with Reduction of SumReduce. Map. This preserves correctness because applying 𝑓 on each partition and then summing is equivalent to summing first and then applying 𝑓 , provided that 𝑓 is linear. (2) Merging Consecutive Map Primitives. Because each Map function applies i… view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Pegasus Syntax. Specifically, our Pegasus Syntax maintains a consistent form with the primitives. In Partition phase, input data and its partitioning rules are explicitly specified. The partitioned data in each segment is used to perform Map operations. In the Map phas…
Figure 7
Figure 7. Figure 7: Impact of per￾flow storage usage on clas￾sification accuracy. 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate Htbot (AUC = 0.8963) Flood (AUC = 0.9986) Cridex (AUC = 0.9993) Virut (AUC = 0.9236) Neris (AUC = 0.9404) Geodo (AUC = 0…
Figure 9
Figure 9. Figure 9: (a–c) Comparison of classification accuracy for different models implemented on the programmable [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages

  1. [1]

    Rishabh Agarwal, Levi Melnick, Nicholas Frosst, Xuezhou Zhang, Ben Lengerich, Rich Caruana, and Geoffrey E Hinton. 2021. Neural additive models: Interpretable machine learning with neural nets.Advances in neural information processing systems34 (2021), 4699–4711

    work page 2021

  2. [2]

    Abd AlRhman AlQiam, Yuanjun Yao, Zhaodong Wang, Satyajeet Singh Ahuja, Ying Zhang, Sanjay G Rao, Bruno Ribeiro, and Mohit Tawar- malani. 2024. Transferable Neural WAN TE for Changing Topologies. InProceedings of the ACM SIGCOMM 2024 Conference. 86–102

    work page 2024

  3. [3]

    Pat Bosshart, Dan Daly, Glen Gibb, Martin Izzard, Nick McKeown, Jennifer Rexford, Cole Schlesinger, Dan Talayco, Amin Vahdat, George Varghese, et al. 2014. P4: Programming protocol-independent packet processors.ACM SIGCOMM Computer Communication Review44, 3 (2014), 87–95

    2014

  4. [4]

    Pat Bosshart, Glen Gibb, Hun-Seok Kim, George Varghese, Nick McK- eown, Martin Izzard, Fernando Mujica, and Mark Horowitz. 2013. For- warding metamorphosis: Fast programmable match-action processing in hardware for SDN.ACM SIGCOMM Computer Communication Review43, 4 (2013), 99–110

    work page 2013

  5. [5]

    Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. 2018. {TVM}: An automated{End-to-End} optimizing compiler for deep learning. In13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18). 578–594

    work page 2018

  6. [6]

    Intel Corporation. 2024. Barefoot Tofino 2. https: //www.intel.cn/content/www/cn/zh/products/details/network-io/ intelligent-fabric-processors/tofino-2.html. (2024)

    work page 2024

  7. [7]

    Intel Corporation. 2024. Barefoot Tofino Series. https: //www.intel.com/content/www/us/en/products/details/network-io/ intelligent-fabric-processors/tofino.html. (2024)

    work page 2024

  8. [8]

    Sajjad Dadkhah, Hassan Mahdikhani, Priscilla Kyei Danso, Alireza Zohourian, Kevin Anh Truong, and Ali A Ghorbani. 2022. Towards the development of a realistic multidimensional IoT profiling dataset. In2022 19th Annual International Conference on Privacy, Security & Trust (PST). IEEE, 1–11

    work page 2022

Show all 68 references
  1. [9]

    Jeff Dean, David Patterson, and Cliff Young. 2018. A new golden age in computer architecture: Empowering the machine-learning revolution. IEEE Micro38, 2 (2018), 21–29

    2018

  2. [10]

    Min Du, Feifei Li, Guineng Zheng, and Vivek Srikumar. 2017. Deeplog: Anomaly detection and diagnosis from system logs through deep learning. InProceedings of the 2017 ACM SIGSAC conference on computer and communications security. 1285–1298

    2017

  3. [11]

    Chuanpu Fu, Qi Li, Meng Shen, and Ke Xu. 2024. Detecting tunneled flooding traffic via deep semantic analysis of packet length patterns. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security. 3659–3673

    2024

  4. [12]

    Massimo Gallo, Alessandro Finamore, Gwendal Simon, and Dario Rossi

  5. [13]

    Gerard Drapper Gil, Arash Habibi Lashkari, Mohammad Mamun, and Ali A Ghorbani. 2016. Characterization of encrypted and VPN traffic using time-related features. InProceedings of the 2nd international conference on information systems security and privacy (ICISSP 2016). SciTePre...

    2016

  6. [14]

    Ian Goodfellow. 2016. Deep learning. (2016)

    2016

  7. [15]

    Dongqi Han, Zhiliang Wang, Wenqi Chen, Kai Wang, Rui Yu, Su Wang, Han Zhang, Zhihua Wang, Minghui Jin, Jiahai Yang, et al. 2023. Anom- aly Detection in the Open World: Normality Shift Detection, Explana- tion, and Adaptation.. InNDSS

    2023

  8. [16]

    Mingcong Han, Hanze Zhang, Rong Chen, and Haibo Chen

  9. [17]

    Jordan Holland, Paul Schmitt, Nick Feamster, and Prateek Mittal. 2021. New directions in automated traffic analysis. InProceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. 3366–3383

    2021

  10. [18]

    Thomas Holterbach, Edgar Costa Molero, Maria Apostolaki, Alberto Dainotti, Stefano Vissicchio, and Laurent Vanbever. 2019. Blink: Fast connectivity recovery entirely in the data plane. In16th USENIX Sym- posium on Networked Systems Design and Implementation (NSDI 19). 161–176

    2019

  11. [19]

    Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. 2018. Quantized neural networks: Training neural networks with low precision weights and activations.Journal of Machine Learning Research18, 187 (2018), 1–30

    2018

  12. [20]

    Broadcom Inc. 2024. Trident 5 / BCM78800 Series. https: //www.broadcom.com/products/ethernet-connectivity/switching/ strataxgs/bcm78800. (2024)

    2024

  13. [21]

    Alon Jacovi, Oren Sar Shalom, and Yoav Goldberg. 2015. Understanding convolutional neural networks for text classification.arXiv preprint arXiv:1809.08037(2015)

    2015 arXiv

  14. [22]

    Syed Usman Jafri, Sanjay Rao, Vishal Shrivastav, and Mohit Tawar- malani. 2024. Leo: Online{ML-based} Traffic Classification at{Multi- Terabit} Line Rate. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). 1573–1591

    2024

  15. [23]

    2019.Deep learning

    John D Kelleher. 2019.Deep learning. MIT press

    2019

  16. [24]

    Alberto Lerner, Davide Zoni, Paolo Costa, and Gianni Antichi. 2024. Rethinking the Switch Architecture for Stateful In-network Computing. InProceedings of the 23rd ACM Workshop on Hot Topics in Networks. 273–281

    2024

  17. [25]

    Chenning Li, Arash Nasr-Esfahany, Kevin Zhao, Kimia Noorbakhsh, Prateesh Goyal, Mohammad Alizadeh, and Thomas E Anderson. 2024. m3: Accurate Flow-Level Performance Estimation using Machine Learning. InProceedings of the ACM SIGCOMM 2024 Conference. 813– 827

    2024

  18. [26]

    Eric Liang, Hang Zhu, Xin Jin, and Ion Stoica. 2019. Neural packet classification. InProceedings of the ACM Special Interest Group on Data Communication. 256–269

    2019

  19. [27]

    Yisroel Mirsky, Tomer Doitshman, Yuval Elovici, and Asaf Shabtai

  20. [28]

    Yarin Perry, Felipe Vieira Frujeri, Chaim Hoch, Srikanth Kandula, Ishai Menache, Michael Schapira, and Aviv Tamar. 2023. A Deep Learning Perspective on Network Routing.arXiv preprint arXiv:2303.00735 (2023)

    2023 arXiv

  21. [29]

    PyTorch. 2024. Post Training Quantization (PTQ) — Torch-TensorRT. https://pytorch.org/TensorRT/tutorials/ptq.html. (2024). Accessed: 2024-09-18

    2024

  22. [30]

    Qiaofeng Qin, Konstantinos Poularakis, Kin K Leung, and Leandros Tassiulas. 2020. Line-speed and scalable intrusion detection at the network edge via federated learning. In2020 IFIP networking conference (Networking). IEEE, 352–360

    2020

  23. [31]

    Babak Rahbarinia, Roberto Perdisci, Andrea Lanzi, and Kang Li. 2013. Peerrush: Mining for unwanted p2p traffic. InDetection of Intrusions and Malware, and Vulnerability Assessment: 10th International Confer- ence, DIMV A 2013, Berlin, Germany, July 18-19, 2013. Proceedings 10....

    2013

  24. [32]

    Michael Seufert, Katharina Dietz, Nikolas Wehner, Stefan Geißler, Joshua Schüler, Manuel Wolz, Andreas Hotho, Pedro Casas, Tobias Hoßfeld, and Anja Feldmann. 2024. Marina: Realizing ML-Driven Real- Time Network Traffic Monitoring at Terabit Scale.IEEE Transactions on Network a...

    2024

  25. [33]

    Vishal Shrivastav. 2022. Programmable multi-dimensional table filters for line rate network functions. InProceedings of the ACM SIGCOMM 2022 Conference. 649–662

    2022

  26. [34]

    Giuseppe Siracusano and Roberto Bifulco. 2018. In-network neural networks.arXiv preprint arXiv:1801.05731(2018)

    2018 arXiv

  27. [35]

    Giuseppe Siracusano, Salvator Galea, Davide Sanvito, Mohammad Malekzadeh, Gianni Antichi, Paolo Costa, Hamed Haddadi, and Roberto Bifulco. 2022. Re-architecting traffic analysis with neural network interface cards. In19th USENIX Symposium on Networked Systems Design and Implem...

    2022

  28. [36]

    Tushar Swamy, Alexander Rucker, Muhammad Shahbaz, Ishan Gaur, and Kunle Olukotun. 2022. Taurus: a data plane architecture for per- packet ML. InProceedings of the 27th ACM International Conference on Architectural Support for Programming Languages and Operating Systems. 1099–1114

    2022

  29. [37]

    Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. 2017. Efficient processing of deep neural networks: A tutorial and survey. Proc. IEEE105, 12 (2017), 2295–2329

    2017

  30. [38]

    Ruming Tang, Zheng Yang, Zeyan Li, Weibin Meng, Haixin Wang, Qi Li, Yongqian Sun, Dan Pei, Tao Wei, Yanfei Xu, et al. 2020. Zerowall: Detecting zero-day web attacks through encoder-decoder recurrent neural networks. InIEEE INFOCOM 2020-IEEE Conference on Computer Communication...

    2020

  31. [39]

    Laurens Van Der Maaten, Eric O Postma, H Jaap Van Den Herik, et al

  32. [40]

    Vladimir Naumovich Vapnik, Vlamimir Vapnik, et al. 1998. Statistical learning theory. (1998)

    1998

  33. [41]

    Ulrike Von Luxburg and Bernhard Schölkopf. 2011. Statistical learning theory: Models, concepts, and results. InHandbook of the History of Logic. Vol. 10. Elsevier, 651–706

    2011

  34. [42]

    Jiazhao Wang, Wenchao Jiang, Ruofeng Liu, Bin Hu, Demin Gao, and Shuai Wang. 2024. {NN-Defined} Modulator: Reconfigurable and Portable Software Modulator on {IoT} Gateways. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). 775–789

    2024

  35. [43]

    Wei Wang, Ming Zhu, Xuewen Zeng, Xiaozhou Ye, and Yiqiang Sheng

  36. [44]

    Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, and Fangxin Wang. 2024. NetLLM: Adapting Large Language Models for Networking. InProceedings of the ACM SIGCOMM 2024 Conference. 661–678

    2024

  37. [45]

    Zhaoqi Xiong and Noa Zilberman. 2019. Do switches dream of machine learning? toward in-network classification. InProceedings of the 18th ACM workshop on hot topics in networks. 25–33

    2019

  38. [46]

    Jinzhu Yan, Haotian Xu, Zhuotao Liu, Qi Li, Ke Xu, Mingwei Xu, and Jianping Wu. 2024. {Brain-on-Switch}: Towards Advanced Intelli- gent Network Data Plane via{NN-Driven} Traffic Analysis at{Line- Speed}. In21st USENIX Symposium on Networked Systems Design and Implementation (N...

    2024

  39. [47]

    Liyan Yang, Yubo Song, Shang Gao, Aiqun Hu, and Bin Xiao. 2022. Griffin: Real-time network intrusion detection system via ensemble of autoencoder in SDN.IEEE Transactions on Network and Service Management19, 3 (2022), 2269–2281

    2022

  40. [48]

    Mingran Yang, Alex Baban, Valery Kugel, Jeff Libby, Scott Mackie, Swamy Sadashivaiah Renu Kananda, Chang-Hong Wu, and Manya Ghobadi. 2022. Using trio: juniper networks’ programmable chipset- for emerging in-network applications. InProceedings of the ACM SIG- COMM 2022 Conferen...

    2022

  41. [49]

    Yucheng Yin, Zinan Lin, Minhao Jin, Giulia Fanti, and Vyas Sekar

  42. [50]

    Zhuolong Yu, Chuheng Hu, Jingfeng Wu, Xiao Sun, Vladimir Braver- man, Mosharaf Chowdhury, Zhenhua Liu, and Xin Jin. 2021. Pro- grammable packet scheduling with a single queue. InProceedings of the 2021 ACM SIGCOMM 2021 Conference. 179–193

    2021

  43. [51]

    Jinxiong Zhang. 2021. Yet Another Representation of Binary Decision Trees: A Mathematical Demonstration.arXiv preprint arXiv:2101.07077 (2021)

    2021 arXiv

  44. [52]

    Ye Zhang and Byron Wallace. 2015. A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification.arXiv preprint arXiv:1510.03820(2015)

    2015 arXiv

  45. [53]

    Yikai Zhao, Wenrui Liu, Fenghao Dong, Tong Yang, Yuanpeng Li, Kaicheng Yang, Zirui Liu, Zhengyi Jia, and Yongqiang Yang. 2023. P4LRU: towards an LRU cache entirely in programmable data plane. InProceedings of the ACM SIGCOMM 2023 Conference. 967–980

    2023

  46. [54]

    Changgang Zheng, Zhaoqi Xiong, Thanh T Bui, Siim Kaupmees, Riyad Bensoussane, Antoine Bernabeu, Shay Vargaftik, Yaniv Ben-Itzhak, and Noa Zilberman. 2022. IIsy: Practical in-network classification. arXiv preprint arXiv:2205.08243(2022)

    2022 arXiv

  47. [55]

    InProceedings of the ACM SIGCOMM 2022 Conference

    Practical gan-based synthetic ip header trace generation using netshare. InProceedings of the ACM SIGCOMM 2022 Conference. 458– 472

    2022

  48. [56]

    Liyan Zheng, Haojie Wang, Jidong Zhai, Muyan Hu, Zixuan Ma, Tuowei Wang, Shuhong Huang, Xupeng Miao, Shizhi Tang, Kezhao Huang, et al . 2023. {EINNET}: Optimizing tensor programs with {Derivation-Based} transformations. In17th USENIX Symposium on Operating Systems Design and I...

    2023

  49. [57]

    Guangmeng Zhou, Xiongwen Guo, Zhuotao Liu, Tong Li, Qi Li, and Ke Xu. 2024. TrafficFormer: An Efficient Pre-trained Model for Traffic Data. In2025 IEEE Symposium on Security and Privacy (SP). IEEE Computer Society, 102–102

    2024

  50. [58]

    Guangmeng Zhou, Zhuotao Liu, Chuanpu Fu, Qi Li, and Ke Xu. 2023. An efficient design of intelligent network data plane. In32nd USENIX Security Symposium (USENIX Security 23). 6203–6220

    2023

  51. [59]

    Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. 2016. Dorefa-net: Training low bitwidth convolu- tional neural networks with low bitwidth gradients.arXiv preprint 14 arXiv:1606.06160(2016)

    2016 arXiv

  52. [60]

    Shitong Zhu, Shasha Li, Zhongjie Wang, Xun Chen, Zhiyun Qian, Srikanth V Krishnamurthy, Kevin S Chan, and Ananthram Swami

  53. [61]

    Changgang Zheng, Mingyuan Zang, Xinpeng Hong, Riyad Bensous- sane, Shay Vargaftik, Yaniv Ben-Itzhak, and Noa Zilberman. 2022. Au- tomating in-network machine learning.arXiv preprint arXiv:2205.08824 (2022)

    2022 arXiv

  54. [67]

    InProceedings of the 16th International Conference on emerging Networking EXperiments and Technologies

    You do (not) belong here: detecting DPI evasion attacks with context learning. InProceedings of the 16th International Conference on emerging Networking EXperiments and Technologies. 183–197

  55. [68]

    Eric R Ziegel. 2003. The elements of statistical learning. (2003). 15

    2003

  56. [2009]

    Dimensionality reduction: A comparative review.Journal of Machine Learning Research10, 66-71 (2009), 13

    2009

  57. [2017]

    In2017 International conference on information networking (ICOIN)

    Malware traffic classification using convolutional neural net- work for representation learning. In2017 International conference on information networking (ICOIN). IEEE, 712–717

  58. [2018]

    Kitsune: an ensemble of autoencoders for online network intru- sion detection.arXiv preprint arXiv:1802.09089(2018). 13

    2018 arXiv

  59. [2020]

    InProceedings of the SIGCOMM’20 Poster and Demo Sessions

    Real-time deep learning based traffic analytics. InProceedings of the SIGCOMM’20 Poster and Demo Sessions. 76–78

  60. [2022]

    In16th USENIX Symposium on Oper- ating Systems Design and Implementation (OSDI 22)

    Microsecond-scale preemption for concurrent {GPU- accelerated}{ DNN} inferences. In16th USENIX Symposium on Oper- ating Systems Design and Implementation (OSDI 22). 539–558

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.