Bohatei:)Flexible)and)Elas2c)) DDoSDefense - USENIX · Bohatei:)Flexible)and)Elas2c)) DDoSDefense...

36
Bohatei: Flexible and Elas2c DDoS Defense Seyed K. Fayaz, Yoshiaki Tobioka, Vyas Sekar, Michael Bailey h5ps://github.com/ddosdefense/bohatei

Transcript of Bohatei:)Flexible)and)Elas2c)) DDoSDefense - USENIX · Bohatei:)Flexible)and)Elas2c)) DDoSDefense...

Bohatei:  Flexible  and  Elas2c    DDoS  Defense    

Seyed  K.  Fayaz,  Yoshiaki  Tobioka,    Vyas  Sekar,  Michael  Bailey  

h5ps://github.com/ddos-­‐defense/bohatei  

DDoS  a5acks  are  geCng  worse

Increasing  in  number  

Threatpost,  7/31/2015  

The  New  York  Times,  3/30/2015  

Increasing  in  volume  Increasing  in  diversity  

Incapsula,  11/12/2014  

2  Arbor  Networks,  2/14/2014   Radware,  10/7/2014  

Cloudflare,  3/27/2013  

Imperva,  2015  

Techworld,  7/16/2014  

High  cost  on  vicIms  

         

Intranet  

DDoS  Defense  Today:    Expensive  Proprietary  Hardware

3  

Assets  

LimitaIon:  Fixed  funcIonality

4  

         

Intranet  

Assets  

What  if  new  types  of  a5acks  emerge?  

LimitaIon:  Fixed  capacity

t1   t2   t3   Ime  

fixed  capacity  

a5ack  vol.(Gbps)  

t4  

5  

waste  waste  

         

Intranet  

Assets  

LimitaIon:  Fixed  locaIon

•  AddiIonal  traffic  latency  due  to  waypoinIng  •  RouIng  hacks  to  enforce  defense  

6  

         

source  

desInaIon  

✗  shortest  path  

         

Need  flexibility  w.r.t.  a5ack  type    

7  

Today:  Hardware  appliance  res.  footprint=240Gbps  Today:  Hardware  appliance  res.  footprint=240Gbps  

Assets  

         

Need  Flexibility  w.r.t  A5ack  LocaIons  

8  

Today:  Hardware  appliance  res.  footprint=240Gbps  Today:  Hardware  appliance  res.  footprint=240Gbps  

Assets  A  

B  

C  

         

Need  ElasIcity  w.r.t.  A5ack  Volume  

9  

Today:  Hardware  appliance  res.  footprint=240Gbps  Today:  Hardware  appliance  res.  footprint=240Gbps  

Assets  

Bohatei  in  a  nutshell..  

A  pracIcal  ISP-­‐scale  system    for  Flexible  and  ElasIc  DDoS  Defense  via  Socware-­‐Defined  Networking  (SDN)  &  Network  FuncIons  VirtualizaIon  (NFV)    à  React  to  500  Gbps  scale  a5acks  in  1  min!  

10  

Outline •  MoIvaIon  •  Background  on  SDN/NFV    •  Bohatei  overview  and  challenges  •  System  design  •  ImplementaIon    •  EvaluaIon  •  Conclusions  

11  

12  

Centralized  management  +  Open  config  APIs  

Controller    

“Flow”   FwdAc2on  …   …  

“Flow”   FwdAc2on  …   …  

“Flow”   FwdAc2on  …   …  

Socware-­‐Defined  Networking  (SDN)  

Network  FuncIons  VirtualizaIon  (NFV)  

13  

Proxy     Firewall   IDS/IPS   AppFilter  Today:  Standalone  and  Specialized  

Commodity  hardware  

Why  are  SDN/NFV  useful  for  DDoS  defense?

14  

Expensive    

Fixed  funcIonality    

Fixed  capacity    

Fixed  locaIon  

NFV  

SDN  

Our  Work:  Bring  these  benefits  to  DDoS  Defense  

Outline •  MoIvaIon  •  Background  on  SDN/NFV    •  Bohatei  overview  and  challenges  •  System  design  •  ImplementaIon    •  EvaluaIon  •  Conclusions  

15  

Bohatei  Vision:    Flexible  +  ElasIc  Defense  via  SDN/NFV

16  

SDN/NFV  Controller

DC2 DC1 customer  intranet  

VM

aJack  traffic

defense  policy  

         

ISP

Bohatei  Controller  Workflow  

17  

Predict  a5ack  pa5ern  

Decide  how  many  VMs,    what  types,  where  

Configure  network  to    route  traffic  

Strategy  layer  

Resource    management  

Network    orchestraIon  

Threat  model:  general,  dynamic  adversaries •  Targets  one  or  more  customers  •  A5acker  has  a  fixed  “budget”  w.r.t.  total  a5ack  volume  

18  

do{ Pick_Target() Pick_Attack_Type() Pick_Attack_Volume() Pick_Attack_Ingress() Observe_and_Adapt()

}

Bohatei  Design  Challenges  

19  

Strategy  layer  

Resource    management  

Network    orchestraIon  

Resilient  to  adaptaIon?  

Fast    algorithms?  

Scalable    SDN?  

Predict  a5ack  pa5ern  

Decide  how  many  VMs,    what  types,  where  

Configure  network  to    route  traffic  

Outline •  MoIvaIon  •  Background  on  SDN/NFV    •  Bohatei  overview  and  challenges  •  System  design  •  ImplementaIon    •  EvaluaIon  •  Conclusions  

20  

Naïve  resource  management  is  too  slow!

21  

Global  opImizaIon  

Takes  hours  to  solve…

Types,  numbers,  and  locaIons  of  VMs?  RouIng  decisions?  

Suspicious  traffic  predicIons Defense  library

Compute/network  resources

Our  Approach:  Hierarchical  +  Greedy

22  

ISP-­‐level  Greedy

… Per  datacenter  N Per  datacenter  1  

How  much  traffic  to  DC1  

Which  VM  slots  in  DC1  

How  much  traffic  to  DCN  

Which  VM  slots  in  DCN  

Suspicious  traffic  predicIons Defense  library

Compute/network  resources

Port1 Port2

Port3

A  reacIve,  per-­‐flow  controller  will  be  a  new  vulnerability

23  

VM1

VM2

SW

Controller

packet1  

VM3

Flow outPort Switch  Forwarding  Table

Flow1 Port  2

Flow100 Port  3

packet100  

ReacIve,  per-­‐flow  isn’t  scalable

… …

Port1

Port  3

Port  2

VM1

VM2

VM3

Idea:  ProacIve  tag-­‐based  steering

24  

Port  2

SW

Controller

Port  3

Context Tag Tag outPort

ProacMve  set  up

ProacIve  per-­‐VM  tagging  enables  scaling

Benign Suspicious

1 2

1 2

2  packet100  

packet1   1  packet1  

packet100  

Dynamic  adversaries  can  game  the  defense  Adversary’s  goals:  

1.  Increase  defense  resource  consumpIon    2.  Succeed  in  delivering  a5ack  traffic    

Simple  predicIon  (e.g.,  prev.  epoch,  avg)  can  be  gamed

t1   t2   t3   Ime  

SYN  flood  predicted  aJack  volume  for  t4  

A5ack  vol.(Gbps)  

t4  

DNS  amp.  

25  

Our  approach:  Online  adaptaIon  •  Metric  of  Success  =  “Regret  minimizaIon”  à    How  worse  than  best  staIc  strategy  in  hindsight?  

 •  Borrow  idea  from  online  algorithms:  Follow  the  perturbed  leader  (FPL)  strategy    

•  IntuiIon:  PredicIon  =  F  (Obs.  History  +  Random  Noise)    •  This  provably  minimizes  the  regret  metric  

26  

PuCng  it  together

PredicIon  strategy    

launching  VMs,  traffic  path  set  up  

predicts  volume  of    suspicious  traffic  of    each  a5ack  type  at    

each  ingress  

OrchestraIon  

quanIty,  type,  locaIon  of  VMs  

suspicious    traffic  spec.  

27  

DC2 DC1 customer  intranet  

VM

aJack  traffic          

ISP

Resource  management  

defense  policy  

Outline •  MoIvaIon  •  Background  on  SDN/NFV    •  Bohatei  overview  and  challenges  •  System  design  •  ImplementaIon    •  EvaluaIon  •  Conclusions  

28  

Defense  policy  library  

29  

Analyze  Srces:  count    

SYN  –  SYN/ACK  per  source  

SYNPROXY  

[LegiImate]  

OK  

LOG   DROP  

[Unknown]  

[A5ack]   [A5ack]  

•  A  defense  graph  per  a5ack  type  •  Customized  interconnecIon  of  defense  modules  •  Open  source  defense  VMs  

Example  (SYN  flood  defense)

[LegiImate]  

ImplementaIon  

30  

FlowTags-­‐enabled    defense  VMs  (e.g.,  Snort)  

OpenDaylight  

resource  manager  

FlowTags  (Fayaz  et  al.,  NSDI’14)  

13  20-­‐core  Intel  Xeon  machines  

OpenFlow  

h5ps://github.com/ddos-­‐defense/bohatei  

KVM  

Control  Plane  

Data  Plane  

defense    library  

Switches  (OVS)  

…  

Outline •  MoIvaIon  •  Background  on  SDN/NFV    •  Bohatei  overview  and  challenges  •  System  design  •  ImplementaIon    •  EvaluaIon  •  Conclusions  

31  

EvaluaIon  quesIons  •  Does  Bohatei  respond  to  a5acks  rapidly?  

•  Can  Bohatei  handle  ≈500  Gbps  a5acks?  

•  Can  Bohatei  successfully  cope  with  dynamic  adversaries?  

 

32  

Responsiveness

33  

Bohatei  restores  performance  of  benign  traffic  ≈  1  min.

•  Hierarchical  resource  management:  – A  few  milliseconds  (vs.  hours)  – OpImality  gap  <  1%  

26

10

0 20 40 60 80 100 120 140

Benig

n tra

ffic

thro

ughput (G

bps)

Time (s)

attack starts

SYN floodDNS amp.

Elephant flowUDP flood

Scalability:  Forwarding  table  size

34  

Per-­‐VM  tagging  cuts  #rules  by  3-­‐4  orders  of  magnitude  ProacIve  setup  reduces  Ime  by  3-­‐4  orders  of  magnitude  

10

1,000

100,000

10e+06

100 200 300 400 500Ma

x re

qu

ire

d n

um

be

ro

f ru

les

on

a s

witc

h

Attack traffic volume (Gbps)

Bohateiper-flow rules

Adversarial  resilience

35  

Bohatei  online  adaptaIon  strategy  minimizes  regret.

0

10

20

30

40

50

60

RandIngress

RandAttack

RandHybrid

SteadyFlipPrevEpoch

Re

gre

t w

.r.t

. vo

lum

eo

f su

cce

ssfu

l att

ack

s (%

)

UniformPrevEpoch

Bohatei

Conclusions

36  

•  DDoS  defense  today  :  Expensive,  Inflexible,  and  InelasIc  

•  Bohatei:  SDN/NFV  for  flexible  and  elasIc  DDoS  defense    

•  Key  Challenges:  Responsiveness,  scalability,  resilience  

•  Main  soluIon  ideas:  – Hierarchical  resource  management  – ProacIve,  tag-­‐based  orchestraIon  – Online  adaptaIon  strategy  

•  Ideas  may  be  applicable  to  other  security  problems  

•  Scalable  +  Can  react  to  very  large  a5acks  quickly!