Inno a i e App oaches o Ne wo k
Analysis and Op imiza ion: Le e aging
Deep Lea ning and P og ammable
Ha dwa e
1s A iel G´
oes de Cas o
Uni e sidade Es adual de Campinas (UNICAMP)
Campinas - SP, B azil
h ps://o cid.o g/0000-0002-5391-5082
2nd Ch is ian Es e e Ro henbe g
Uni e sidade Es adual de Campinas (UNICAMP)
Campinas - SP, B azil
h ps://o cid.o g/0000-0003-3109-4305
Abs ac —Ne wo k demand o eal- ime applica-
ions like sel -d i ing ca s and cloud gaming s ains
exis ing ne wo ks. La ency and conges ion hu use
expe ience. Realis ic es ing is i al o imp o ing ne -
wo ks, bu eal-wo ld da a is sca ce. In his con ex ,
we p opose o analyze exis ing ne wo k da a and
iden i y a ic pa e ns and anomalies. We belie e his
knowledge can be used o eed gene a i e ad e sa ial
ne wo k (GAN) models, which can c ea e ealis ic
syn he ic da a, supplemen ing exis ing eal aces
while p o ec ing end-use p i acy. This augmen ed
da a can hen be used, o ins ance, o empowe
imp o ed ou ing algo i hms designed o bene i om
p og ammable ha dwa e (e.g., Sma NICs) and col-
lec ed da a plane me ics, pa ing he way o imp o ed
ne wo k pe o mance and enhanced use expe ience
wi h mo e au onomous decisions. This pape p esen s
ou ini ial analysis o syn he ic ne wo k da a gen-
e a ion echnologies and summa izes he main ideas
guiding my Ph. D. esea ch.
Index Te ms—p og ammable ha dwa e, neu al ne -
wo ks, and ne wo k ace gene a ion.
I. INTRODUCTION
The su ge in ne wo k demand ueled by eal-
ime, da a-in ensi e applica ions like heal hca e, au-
onomous ehicles, and cloud gaming has in ensi-
ied he need o enhanced ne wo k pe o mance.
Howe e , p oduc ion ne wo ks o en su e om
issues like la ency and conges ion, impac ing use
expe ience and se ice quali y. Also, ob aining eal
ne wo k da a o hese asks is o en challenging
o se e al easons. Fi s , ne wo k da a may con ain
sensi i e o p i a e in o ma ion o he use s o
o ganiza ions, which aises e hical and legal issues
o sha ing o publishing hem. Second, ne wo k
This wo k was pa ially suppo ed by he Inno a ion Cen-
e , E icsson S.A., and by he Sao Paulo Resea ch Founda-
ion (FAPESP), g an 2021/00199-8, CPE SMARTNESS.
Disc imina o
Gene a ed
samples
Real
samples
Gene a o
Real?
Fake?
Loss
Loss
P edic ions
E o adjus men
E o adjus men
Fig. 1. GAN a chi ec u e o e iew.
da a may be sca ce o ou da ed [1], especially o
eme ging o e ol ing ne wo k scena ios (e.g., 5G,
IoT, SDN). Thi d, ne wo k da a may be biased o
incomple e [2], [3], which limi s he gene aliza ion
and obus ness o he ne wo k analysis models.
To e ec i ely de elop new algo i hms and solu-
ions ha ca e o he di e se equi emen s o such
applica ions unde ealis ic condi ions, i is impe a-
i e o es hem in en i onmen s ha closely esem-
ble eal-wo ld scena ios. I necessi a es using eal
de ices and inco po a ing communica ion logs. To
o e come he challenges men ioned abo e, ne wo k
da a gene a ion echniques ha e been p oposed o
c ea e syn he ic o ealis ic ne wo k da a ha can be
used o ne wo k analysis asks. These echniques
mimic he cha ac e is ics and beha io s o eal ne -
wo k da a, such as packe heade s, payloads, lows,
p o ocols, and a ic pa e ns (e.g., ideo s eaming,
VoIP). Ne wo k da a gene a ion echniques can also
in oduce a ia ions and anomalies o he ne wo k
da a o simula e di e en ne wo k condi ions and
scena ios.
Ne wo k da a gene a ion echniques can be clas-
si ied in o wo main ca ego ies: model-based and
2024 IEEE 10 h In e na ional Con e ence on Ne wo k So wa iza ion (Ne So )
979-8-3503-6958-8/24/$31.00 ©2024 IEEE 239
2024 IEEE 10 h In e na ional Con e ence on Ne wo k So wa iza ion (Ne So ) | 979-8-3503-6958-8/24/$31.00 ©2024 IEEE | DOI: 10.1109/Ne So 60951.2024.10588944
Au ho ized licensed use limi ed o: KTH Royal Ins i u e o Technology. Downloaded on Augus 27,2025 a 11:24:25 UTC om IEEE Xplo e. Res ic ions apply.
ace-based. Model-based [4], [5] echniques use
ma hema ical models o s a is ical me hods o gen-
e a e ne wo k da a om sc a ch. Model-based ech-
niques can cap u e eal ne wo k da a’s gene al
p ope ies and dis ibu ions, such as packe in e -
a i al imes, packe sizes, low du a ion, and low
a e. Model-based echniques can also inco po a e
di e en ne wo k pa ame e s and con igu a ions,
such as he numbe o hos s, connec ions, a ic
ypes, and a ic olumes. Howe e , model-based
echniques may no be able o ep oduce he speci ic
ea u es and dynamics o eal ne wo k da a, such
as p o ocol-speci ic ields, applica ion-speci ic con-
en s, and empo al o spa ial co ela ions. On he
o he hand, ace-based echniques [6] use exis ing
ne wo k aces o PCAPs as inpu s o gene a e
new ne wo k da a. I can p ese e he ealis ic
and de ailed aspec s o eal ne wo k da a, such
as packe heade s, payloads, lows, p o ocols, and
a ic pa e ns. Wi h his new pa adigm, we can
modi y o manipula e he exis ing ne wo k aces
o PCAPs o c ea e new ne wo k da a wi h di e -
en cha ac e is ics o beha io s while anonymizing
sensi i e o p i a e in o ma ion (e.g., end-use IPs)
in he ne wo k aces.
A p omising ace-based app oach lies in ha ness-
ing he ecen s ides made in Gene a i e Ad e -
sa ial Ne wo ks (GANs). Jus as GANs [7] ha e
e olu ionized he gene a ion o high-quali y im-
ages, hey hold he po en ial o eshape he land-
scape o PCAP (packe cap u e) gene a ion. Fig-
u e 1 summa izes a GAN a chi ec u e o e iew.
I encompasses wo dis inc Neu al Ne wo ks —
an en i y designa ed as he gene a o and ano he
as he disc imina o — ha engage in an in e play
ollowing he p inciples o game heo y, as delin-
ea ed in [8]. The undamen al ope a ional pa adigm
in ol es he gene a o ne wo k p oducing syn he ic
da a samples o decei e he disc imina o . In pa al-
lel, he disc imina o ne wo k unde akes he ole
o a judge, assessing he simila i y be ween eal
and gene a ed/syn he ic da a. The main objec i e
is o c ea e a scena io whe ein he disc imina o ’s
capaci y o disce n ac ual da a om i s syn he ic
coun e pa s is ma kedly diminished.
When employed as a syn he ic da a gene a o ,
GANs wo k as a simula o . Wi hin his con ex ,
he syn he ic da a p oduced emula es he inhe en
dis ibu ion o he o iginal da ase [9], he eby en-
su ing he p ese a ion o p i acy conside a ions.
Mo eo e , hese ne wo ks p o e ins umen al in
asks such as da ase augmen a ion and balancing,
culmina ing in a da ase cha ac e ized by enhanced
ep esen a ional capaci y. Consequen ly, he esul-
an model eme ges as a condui o sha e in ica e
dynamics o eal en i onmen s while ob usca ing
inhe en complexi ies and main aining da a quali y
in eg i y.
The es o he pape is o ganized as ollows.
Sec ion II summa izes he s a e o he a on syn-
he ic a ic gene a ion. Sec ion III p esen s he
me hodology we hope o use o answe he main
ques ions iden i ied. Finally, Sec ion IV concludes
he a icle, ecapi ula ing he opics co e ed and
illus a ing he u u e s eps o ou esea ch.
II. STATE OF THE ART
Recen ly, ealis ic a ic gene a ion elied p ima -
ily on neu al ne wo k a chi ec u es such as Gene -
a i e Ad e sa ial Ne wo ks (GANs) [6], [9], [10]
o di usion models [11]. Fo ins ance, he GAN
a chi ec u e wo ks as ollows: a gene a o p oduces
syn he ic da a ha mimics he eal da a dis ibu ion,
while he disc imina o ies o dis inguish be ween
eal and syn he ic da a. Bo h a e ained ad e -
sa ially un il hey each an equilib ium whe e he
disc imina o canno di e en ia e (i.e., disc imina e)
be ween eal and syn he ic da a. Simila ly, di usion
models ha e a con olled and g adual aining p o-
cess bu a e mo e compu a ionally expensi e.
PcapGAN [6] p oposes a me hod o gene a ing
ealis ic PCAP iles ha p ese e he s yle and
s uc u e o eal PCAP iles. The echnique uses a
s yle-based GAN a chi ec u e ha can con ol he
s yle o he gene a ed packe s a di e en le els o
abs ac ion. SIP-GAN [10] in oduces a me hod o
gene a ing ealis ic SIP (Session Ini ia ion P o ocol)
a ic ha can be used o es ing VoIP (Voice o e
IP) sys ems. The echnique uses a condi ional GAN
(cGAN) a chi ec u e ha can gene a e SIP messages
wi h di e en ypes, such as INVITE, ACK, BYE,
CANCEL, and OPTIONS. Ne Di usion [11] le e -
ages a ixed-leng h packe ep esen a ion [12] o
ans o m packe s in o images ha can be easily ma-
nipula ed. Despi e ha , i s p o ocol ule-compliance
app oach is pos -gene a i e, which means he gen-
e a ed packe is no p o ocol-complian and mus be
modi ied o e lec he desi ed ou pu .
While he men ioned wo k has demons a ed he
e icacy o neu al ne wo ks in ne wo k analysis
asks, i is impo an o no e hey ha e p ima ily
been execu ed on gene ic-pu pose CPUs. How-
e e , he landscape o ne wo k ha dwa e is apidly
e ol ing, p esen ing new oppo uni ies o le e age
specialized ha dwa e o enhanced pe o mance and
e iciency. One such p omising a enue is he in-
eg a ion o neu al ne wo ks wi h p og ammable
2024 IEEE 10 h In e na ional Con e ence on Ne wo k So wa iza ion (Ne So )
240
Au ho ized licensed use limi ed o: KTH Royal Ins i u e o Technology. Downloaded on Augus 27,2025 a 11:24:25 UTC om IEEE Xplo e. Res ic ions apply.
ne wo k ha dwa e, such as Sma NICs, o ackle a
di e se ange o ne wo k challenges.
By o loading neu al ne wo k p ocessing asks,
signi ican pe o mance imp o emen s can be
achie ed [13], leading o as e decision-making
and educed la ency. I is pa icula ly ad an ageous
o eal- ime applica ions such as in usion and
anomaly de ec ion, whe e imely esponses a e c i -
ical. Fu he mo e, p og ammable da a plane ha d-
wa e p o ides access o low-le el ne wo k da a
wi h minimal o e head, enabling neu al ne wo ks
o ope a e di ec ly on aw packe s eams wi h local
decisions. Fo ins ance, FcNN [14] is a dis ibu-
i e da a-cen ic compu ing amewo k o econ ig-
u able Sma NIC-based sys ems. I allows he com-
ple e de aching o NN ke nel execu ion con ol logic
sys em scheduling and ne wo k communica ion o
he Sma NICs. I boos s pe o mance by a oiding
con ol dependency wi h CPUs o a ious neu al
ne wo k ke nels and applica ions, including DNNs
and GNNs.
III. OBJECTIVES, RESEARCH QUESTIONS AND
METHODOLOGY
The p oposed esea ch aims o add ess gaps in
he s a e-o - he-a by explo ing he easibili y o
gene a ing ealis ic ne wo k aces wi h di e en
gene a i e app oaches.
A. Resea ch Ques ions
The p oposed esea ch plan deals wi h he ol-
lowing esea ch ques ions.
1) How can we handle speci ic challenges like
sequence gene a ion [6] and empo al de-
pendencies [11], in he con ex o ne wo k
ace gene a ion? Can hose es ic ions be
embedded in o he model, o mus hey be
ea ed in a pos -gene a ion manne ?
2) Is i possible o c ea e a anspa en a ic
gene a ion ool o he end use ?
3) Should we o load ne wo k applica ions (e.g.,
ou ing) wi h he aid o neu al ne wo k
models in o p og ammable ne wo k ha dwa e
(e.g., FPGAs, Sma NICs)? I so, wha a e he
bes o loading s a egies (i.e., hyb id o o al
o loading) o each applica ion?
4) Wha po en ial challenges will i p esen (e.g.,
ene gy consump ion, memo y limi a ion) o
di e en ha dwa e a chi ec u es (e.g., Sma -
NICs, FPGAs)?
B. Wo k Plan
This sec ion p o ides an o e iew o he wo k
plan s ages ha will di ec he esea ch ac i i ies o
his Ph.D. The esul s will be published acco dingly
in all o hese phases. To a ain he objec i es
men ioned ea lie and esea ch ques ions, we will
use he ollowing me hodology.
Add essing Challenges in Ne wo k T ace Gene -
a ion: The i s s ep owa ds an in elligen sys em
ha gene a es ealis ic packe aces is o (i) ga he
a di e se se o eal-wo ld ne wo k aces (PCAP
iles) ep esen ing a ious communica ion scena -
ios and p o ocols and (ii) p ep ocess he collec ed
da ase o ex ac ele an ea u es and no malize
da a o ensu e consis ency ac oss di e en aces.
To do ha , we in end o le e age nP in [12] packe
ep esen a ion. I p o ides a s anda dized bi -le el
ep esen a ion o e e y ne wo k packe , ensu ing
all po en ial heade ields (e en i no p esen in
he o iginal packe ). Fo ins ance, while a TCP
packe will no ha e UDP heade bi s, he nP in
s ill includes placeholde s o hese bi s, ensu ing a
uni o m inpu s uc u e o ML models. Rega ding
he da ase s, Kaggle is a p omising online da a sci-
ence pla o m ha could se e as a s a ing poin o
ga he ing eely a ailable PCAPs. On his websi e,
we ound a p omising da ase wi h a ound 7GB o
ne wo k aces spli in he e ogeneous applica ions
such as Skype and Amazon. Ano he op ion would
be o explo e code pla o ms such as Gi Hub and
Gi Lab o o he pla o ms such as Pape swi hcode,
which g oups pape s and hei espec i e codes –
i.e. i hey a e open sou ce. Howe e , i is ye o be
known whe he he a ailable da a may be biased,
wi h a lo o epea ed in o ma ion ha does no help
he neu al model o lea n di e en a ic scena ios.
Fo example, i we conside a da ase abou bank
aud and conside wo classes ( ue, alse) whe e
mos o he labels a e ” alse”, hen he ained model
may become biased, lea ning/specializing mo e in
ce ain cha ac e is ics ha do no acili a e he iden-
i ica ion o ue posi i es o aud. The same idea
applies o he compu e ne wo k con ex . An ideal
da ase dis ibu ion should be able o ep esen he
en i e – i.e., a leas mos o – da a dis ibu ion o e
ime.
An end-use a ic gene a o ool: Besides ha ing
da a om each applica ion, i is necessa y o c ea e
a neu al ne wo k ha unde s ands he demands o
each applica ion o gene a e ne wo k a ic ha
cap u es he nuances o each eques . Ini ially, he
mos p omising a chi ec u e es ed o c ea e images
mo e ai h ul o he o iginals is Va ia ional Au-
oencode s (VAEs) [15]. Howe e , he e a e some
limi a ions. Fi s , we mus ain he model o line
and e-execu e he PCAPs in he in as uc u e.
2024 IEEE 10 h In e na ional Con e ence on Ne wo k So wa iza ion (Ne So )
241
Au ho ized licensed use limi ed o: KTH Royal Ins i u e o Technology. Downloaded on Augus 27,2025 a 11:24:25 UTC om IEEE Xplo e. Res ic ions apply.
Second, he use has li le con ol o e he ype
o applica ion o be gene a ed o wha ypes o
a ic he ne wo k can gene a e. To achie e his, we
could in eg a e he idea o c ea ing o modi ying an
exis ing la ge language model (LLM) [16] capable
o unde s anding high-le el use eques s in he
ne wo k con ex and gene a ing ne wo k a ic wi h
desi ed cha ac e is ics. Fo example, ideally, we
should be able o gene a e a ic as ollows (o sim-
ila ly) “Ne lix a ic, 1GBps, la ency 50ms, ji e
5ms”. In he p e ious example, he use would no
(ideally) need o wo y abou how hei da a is being
gene a ed, whe e he solu ion’s in e nal p ocesses
would be anspa en o hem. Howe e , he use
could be gua an eed ha he da a dis ibu ion he
eques ed would be gene a ed, in he same way as
he a ic es ic ions (e.g. ji e , la ency).
O loading Ne wo k Applica ions o
P og ammable Ha dwa e: A his poin , we
will conduc a comp ehensi e e iew o a ailable
p og ammable ne wo k ha dwa e op ions, including
Sma NICs and FPGAs ha can be le e aged,
conside ing hei speci ica ions, capabili ies, and
p og ammabili y ea u es o de e mine sui abili y
o o loading ne wo k applica ions. Then, we will
iden i y he mos p ominen ne wo k applica ions
sui able o o loading (i.e., pa ially o e en
o ally) and adap he selec ed applica ions o
execu ion on p og ammable ha dwa e, op imizing
o pe o mance and esou ce u iliza ion.
Analysis o Ha dwa e Limi a ions and T ade-
o s: Conside ing he acquisi ion o he necessa y
ha dwa e, p o iling es s will be ca ied ou on
he a ailable pla o ms and Measu e key pe o -
mance me ics such as p ocessing speed, memo y
bandwid h, and ene gy consump ion unde a ying
wo kloads o pa ially/ o ally o loaded applica-
ions, conside ing ac o s like da a ans e a e and
p o ocol o e head, de e mining he easibili y and
p ac ical implica ions o o loading ne wo k appli-
ca ions o p og ammable ha dwa e in eal-wo ld
deploymen scena ios.
IV. CONCLUSION
This a icle p esen s he main esea ch di ec ions
o my i s -yea Ph.D. In his con ex , he idea o
a ic gene a ion can s ill be widely explo ed o
di e en p o ocols and applica ions, acili a ing he
gene a ion o a ic o he end use anspa en ly.
Fu he mo e, he e is he possibili y o o loading
di e en applica ions on o p og ammable ha dwa e.
We belie e we could bene i om his equipmen
and accele a e he a ic gene a ion p ocess.
REFERENCES
[1] H. N. Qu eshi, U. Masood, M. Manalas as, S. M. A. Zaidi,
H. Fa ooq, J. Fo gea , M. Bou on, S. Bo he, P. Ka ls-
son, A. Rizwan, e al., “Towa ds add essing aining da a
sca ci y challenge in eme ging adio access ne wo ks: A
su ey and amewo k,” IEEE Communica ions Su eys &
Tu o ials, 2023.
[2] T. B¨
uhle , R. Schmid, S. Lu z, and L. Vanbe e , “Gen-
e a ing ep esen a i e, li e ne wo k a ic ou o millions
o code eposi o ies,” in P oceedings o he 21s ACM
Wo kshop on Ho Topics in Ne wo ks, pp. 1–7, 2022.
[3] E. Dai and S. Wang, “Say no o he disc imina ion:
Lea ning ai g aph neu al ne wo ks wi h limi ed sensi i e
a ibu e in o ma ion,” in P oceedings o he 14 h ACM
In e na ional Con e ence on Web Sea ch and Da a Mining,
pp. 680–688, 2021.
[4] G. Ca nei o, “Ns-3: Ne wo k simula o 3,” in UTM lab
mee ing Ap il, ol. 20, pp. 4–5, 2010.
[5] Y. Yin, Z. Lin, M. Jin, G. Fan i, and V. Seka , “P ac i-
cal gan-based syn he ic ip heade ace gene a ion using
ne sha e,” in P oceedings o he ACM SIGCOMM 2022
Con e ence, pp. 458–472, 2022.
[6] B. Dowoo, Y. Jung, and C. Choi, “Pcapgan: Packe cap u e
ile gene a o by s yle-based gene a i e ad e sa ial ne -
wo ks,” in 2019 18 h IEEE In e na ional Con e ence On
Machine Lea ning And Applica ions (ICMLA), pp. 1149–
1154, IEEE, 2019.
[7] I. Good ellow, J. Pouge -Abadie, M. Mi za, B. Xu,
D. Wa de-Fa ley, S. Ozai , A. Cou ille, and Y. Bengio,
“Gene a i e ad e sa ial ne wo ks,” Communica ions o he
ACM, ol. 63, no. 11, pp. 139–144, 2020.
[8] I. Good ellow, J. Pouge -Abadie, M. Mi za, B. Xu,
D. Wa de-Fa ley, S. Ozai , A. Cou ille, and Y. Bengio,
“Gene a i e ad e sa ial ne s,” Ad ances in neu al in o -
ma ion p ocessing sys ems, ol. 27, 2014.
[9] Y. Yin, Z. Lin, M. Jin, G. Fan i, and V. Seka , “P ac-
ical gan-based syn he ic ip heade ace gene a ion us-
ing ne sha e,” in P oceedings o he ACM SIGCOMM
2022 Con e ence, SIGCOMM ’22, (New Yo k, NY, USA),
p. 458–472, Associa ion o Compu ing Machine y, 2022.
[10] A. Meddahi, H. D i a, and A. Meddahi, “Sip-gan: Gen-
e a i e ad e sa ial ne wo ks o sip a ic gene a ion,” in
2021 In e na ional Symposium on Ne wo ks, Compu e s
and Communica ions (ISNCC), pp. 1–6, IEEE, 2021.
[11] X. Jiang, S. Liu, A. Gembe -Jacobson, A. N. Bhagoji,
P. Schmi , F. B onzino, and N. Feams e , “Ne di usion:
Ne wo k da a augmen a ion h ough p o ocol-cons ained
a ic gene a ion,” a Xi p ep in a Xi :2310.08543, 2023.
[12] J. Holland, P. Schmi , N. Feams e , and P. Mi al, “New
di ec ions in au oma ed a ic analysis,” CCS ’21, (New
Yo k, NY, USA), p. 3366–3383, Associa ion o Compu -
ing Machine y, 2021.
[13] C. Campolo, A. Ie a, and A. Molina o, “Ne wo k o
dis ibu ed in elligence: a su ey and u u e pe spec i es,”
IEEE Access, 2023.
[14] A. Guo, T. Geng, Y. Zhang, P. Haghi, C. Wu, C. Tan,
Y. Lin, A. Li, and M. He bo d , “Fcsn: A pga-
cen ic sma nic amewo k o neu al ne wo ks,” in 2022
IEEE 30 h Annual In e na ional Symposium on Field-
P og ammable Cus om Compu ing Machines (FCCM),
pp. 1–2, IEEE, 2022.
[15] C. Doe sch, “Tu o ial on a ia ional au oencode s,” a Xi
p ep in a Xi :1606.05908, 2016.
[16] Y. Chang, X. Wang, J. Wang, Y. Wu, L. Yang, K. Zhu,
H. Chen, X. Yi, C. Wang, Y. Wang, e al., “A su ey on
e alua ion o la ge language models,” ACM T ansac ions
on In elligen Sys ems and Technology, ol. 15, no. 3,
pp. 1–45, 2024.
2024 IEEE 10 h In e na ional Con e ence on Ne wo k So wa iza ion (Ne So )
242
Au ho ized licensed use limi ed o: KTH Royal Ins i u e o Technology. Downloaded on Augus 27,2025 a 11:24:25 UTC om IEEE Xplo e. Res ic ions apply.
