Hash-Based IP Traceback

Alex C. Snoeren†, Craig Partridge,

Luis A. Sanchez, Christine E. Jones, Fabrice Tchakountio,

Stephen T. Kent, W. Timothy Strayer

BBN Technologies†MIT Laboratory for Computer Science

Network Security Risks

• Tools readily available to attackers network server attacks performance degradation attacks

• DOS• DDOS

Single packet attacks (Stop 0A in TCPIP.sys, Teardrop, Ping-of-death)

• Accidental (unintentional) attacks

Approaches

• Firewalls - prevent attack packets from reaching the victim some attack packets look quite innocent hard to predict all possible attacks does not get at the source of the problem continue to consume network resources

• Traceback - identify the source of attack packets For a given packet, find the path to source

Why Traceback is hard

• Internet Protocol permits anonymity Attackers can “spoof” source address

• Fraggle/Smurf, etc

IP forwarding maintains no audit trails

• Some spoofing is legitimate (NATs, mobile IP, etc)

• Attacks may be short-lived• Packets change hop by hop• Routing instability

Why Traceback is hard (continued)

• Network may carry multiple identical packets (attacks, multicast, broadcast)

• Routers may be compromised

• Attackers may be aware they are being traced

• Increasing packet size is frowned on

• Will consume network resources

• Ingress filtering of limited value

Traceback Goal

• Reconstruct the attack path of a packet where the path consists of every router on the path from the source to the victim

• Reconstruct the attack graph which may result from multiple copies of an attack packet injected by different sources

• Need to be able to detect false positives with a high degree of accuracy

Approaches to Traceback

• Path data can be noted in several places In the packet itself [Savage et al.], At the destination [I-Trace], or In the network infrastructure

• Logging: a naïve in-network approach Record each packet forwarding event Can trace a single packet to a source router,

ingress point, or subverted router(s)

Log-Based Traceback

V

R

R1 R2

R3

RR

RR

R4

A R

RR7

R6R5

Challenges to Logging

• Attack path reconstruction is difficult Packet may be transformed as it moves

through the network

• Full packet storage is problematic Memory requirements are prohibitive at

high line speeds (OC-192 is ~10Mpkt/sec)

• Extensive packet logs are a privacy risk Traffic repositories may aid eavesdroppers

Solution: Packet Digesting

• Record only invariant packet content Mask dynamic fields (TTL, checksum, etc.) Store information required to invert packet

transformations at performing router

• Compute packet digests instead Use hash function to compute small digest Store probabilistically in Bloom filters

• Impossible to retrieve stored packets

Invariant Content

Total Length

Identification

Checksum

Ver TOSHLen

TTL Protocol

Source Address

Destination Address

Fragment OffsetMF

DF

Options

Remainder of Payload

First 8 bytes of Payload

28bytes

Impact of Traffic Diversity

1e-06

1e-05

0.0001

0.001

0.01

0.1

1

20 22 24 26 28 30 32 34 36 38 40

Fra

ctio

n of

Col

lided

Pac

kets

Prefix Length (in bytes)

WAN (6031 hp)LAN (2879 hp)

• Variable capacity Easy to adjust Page when full

Bloom Filters

• Fixed structure size Uses 2n bit array Initialized to zeros

• Insertion is easy Use n-bit digest as

indices into bit array

1n bits

2n

bits

H(P)H2(P)

Hk(P)

H3(P)

H1(P)

1

1

1

. . .

Mitigate collisions by using multiple digests

Mistake Propagation is Limited

• Bloom filters may be mistaken Mistake frequency can be controlled Depends on capacity of full filters

• Neighboring routers won’t be fooled Vary hash functions used in Bloom filters Each router select hashes independently

• Long chains of mistakes highly unlikely Probability drops exponentially with length

Adjusting Graph Accuracy

• False positives rate depends on: Length of the attack path Complexity of network topology Capacity of Bloom filters

• Bloom filter capacity is easy to adjust Required filter capacity varies with router

speed and number of neighbors Appropriate capacity settings achieve

linear error growth with path length

Simulation Results

0

0.2

0.4

0.6

0.8

1

0 5 10 15 20 25 30

Length of Attack Path (in hops)

Random Graph

0

0.2

0.4

0.6

0.8

1

0 5 10 15 20 25 30

Length of Attack Path (in hops)

0

0.2

0.4

0.6

0.8

1

0 5 10 15 20 25 30

Length of Attack Path (in hops)

Real ISP, 100% Utilization

0

0.2

0.4

0.6

0.8

1

0 5 10 15 20 25 30

Length of Attack Path (in hops)

Degree-Independent

Exp

ecte

d N

umbe

r of

Fal

se P

ositi

ves Real ISP, Actual Utilization

How long can digests last?

• Filters require 0.5% of link capacity Four OC-3s require 47MB per minute A single drive can store a whole day

• Access times are equally important Current drives can write >3GB per minute OC-192 needs SRAM access times

• Still viable tomorrow 128 OC-192 links need <100GB per minute

Prototype Implementation

• Implemented on a FreeBSD PC router Packet digesting on kernel forwarding path Bloom filters stored in kernel space Zero-copy kernel/user table move

• User-level query-support daemons Supports topology discovery through gated Queries automatically triggered by IDS

Summary

• Hash-based traceback is viable With reasonable memory constraints Supports common packet transforms Timely tracing of individual packets

• Publicly Available Implementation FreeBSD version will be available soon Linux port coming shortly thereafter….

http://www.ir.bbn.com/projects/SPIE