Information about http://www.cs.cmu.edu/~haowen/key-infection.pdf
Key Infection: Smart Trust for Smart…
Tags: adrian perrig, appli, battery energy, building safety, carnegie mellon university, cations, cmu, distribution problem, dynamic network, effi, munication, pect, resource discovery, ross anderson, sensor network, sensor networks, smart dust, sor, storage resources, university of cambridge,Pages: 10
Language: english
Display cached document
Page 1
Page 2
Page 3
Page 4
Page 5
Page 6
Page 7
Page 8
Page 9
Page 10
Key Infection: Smart Trust for Smart Dust
Ross Anderson Haowen Chan Adrian Perrig
University of Cambridge Carnegie Mellon University Carnegie Mellon University
Ross.Anderson@cl.cam.ac.uk haowenchan@cmu.edu perrig@cmu.edu
Abstract and building safety. As sensor networks become cheaper
and more commoditised, they will become attractive to
Future distributed systems may include large self- home users and small businesses, and for other new appli-
organizing networks of locally communicating sen- cations.
sor nodes, any small number of which may be sub- A typical sensor network consists of a large number of
verted by an adversary. Providing security for these small, low-cost nodes that use wireless peer-to-peer com-
sensor networks is important, but the problem is compli- munication to form a self-organized network. They use
cated by the fact that managing cryptographic key ma- multi-hop routing algorithms based on dynamic network
terial is hard: low-cost nodes are neither tamper-proof and resource discovery protocols. To keep costs down and
nor capable of performing public key cryptography effi- to deal with limited battery energy, nodes have fairly min-
ciently. imal computation, communication, and storage resources.
In this paper, we show how the key distribution problem They do not have tamper-proof hardware. We can thus ex-
can be dealt with in environments with a partially present, pect that some small fraction of nodes in a network may be
passive adversary: a node wishing to communicate securely compromised by an adversary over time.
with other nodes simply generates a symmetric key and An interesting example of a sensor network technology
sends it in the clear to its neighbours. Despite the apparent is given by the `Smart Dust' project which is developing
insecurity of this primitive, we can use mechanisms for key tiny sensors [9]. Its goal is to make sensors so small and
updating, multipath secrecy amplification and multihop key cheap that they can be distributed in large numbers over an
propagation to build up extremely resilient trust networks area by random scattering.
where at most a fixed proportion of communications links The security of sensor networks may be important, es-
can be eavesdropped. We discuss applications in which this pecially if they are deployed in military applications or in
assumption is sensible. safety-critical applications such as medical monitoring. In
Many systems must perforce cope with principals who addition to physical destruction or barrage jamming, a range
are authenticated weakly, if at all; the resulting issues have of more subtle attacks may be attempted by a capable, moti-
often been left in the `too hard' tray. One particular inter- vated opponent. The opponent may be passive, simply mon-
est of sensor networks is that they present a sufficiently com- itoring the sensor data flows in order to determine the ex-
pact and tractable version of this problem. We can perform tent and capability of the network; she may be active, and
quantitative analyses and simulations of alternative strate- transmit deceptive messages; she may try selective jamming
gies, some of which we present here. We also hope that this or network flooding; and she may subvert a number of the
paper may start to challenge the common belief that authen- nodes and use them for various active and Byzantine at-
tication is substantially about bootstrapping trust. We argue tacks. The results of such attacks may include the loss of
that, in distributed systems where the opponent can subvert personal privacy, the loss of service of critical sensor sys-
any small proportion of nodes, it is more economic to in- tems, and the consequences of people or equipment taking
vest in resilience than in bootstrapping. erroneous action based on maliciously falsified sensor data.
When designing security for sensor networks, an impor-
tant problem is key distribution the problem of estab-
lishing shared secret keys between sensor nodes. Efficient
1. Introduction key distribution for sensor networks is still an important re-
search area. Devices with capable processors can use well-
Wireless sensor networks are becoming increasingly im- established asymmetric cryptographic techniques, based on
portant for a wide variety of applications such as factory in- Diffie-Hellman key agreement or RSA-based key establish-
strumentation, climate control, environmental monitoring, ment [4, 15]. However, the cheap processors used in sen-
sor nodes often lack the resources to perform asymmetric In the world of wireless sensor networks, such an adver-
cryptography. These devices are expected to cost at most sary is often not realistic. In many applications, we will be
a few tens of cents, so asymmetric cryptography may of- able to assume that she can monitor only some small pro-
ten require too much computation or too much memory. portion of the traffic during initial deployment.
Even where it is feasible, the fundamental limit on sensor We will discuss a number of scenarios in detail in section
node performance is battery energy; each available micro- III; as an example meanwhile, consider a burglar alarm sen-
joule must be carefully apportioned between sensing, com- sor network deployed in an office building containing a few
putation and communication. Digital signatures might thus hundred nodes, each with an effective radio range of 10 me-
be used by an attacker to perform battery-draining denial- tres. If the opponent has already installed so many surveil-
of-service attacks. lance devices in the building that she can monitor all radio
In response to this problem, key-distribution schemes exchanges between nodes, then the building owner is try-
based on symmetric cryptography have been proposed [6, ing to solve the wrong problem! As another example, con-
3]. Typical schemes rely on having each node pre-load a sider a canister of 10,000 smart dust motes deployed via an
large number of keys to have a reasonable probability of artillery shell into enemy territory, and which then commu-
sharing one with a neighbour. However, such schemes still nicate with each other using low-power radio on a single
require a large amount of memory, as well as an infras- channel. Each enemy sensor will only be able to pick up the
tructure to load the keys into the sensor nodes, which may nearest signal; distant signals will interfere with each other.
be too cumbersome for some applications. Other schemes Thus, if our `white dust' is deployed at a much greater den-
set up pairwise keys by physical contact between devices; sity than the enemy's `black dust' defensive motes, most
for example, a number of burglar or fire alarm sensors are of our communications will go unmonitored at least ini-
each touched briefly against a controller before emplace- tially. (Of course, once the presence of white dust has been
ment [17]. Since each node must be physically brought into reported, several canisters of black dust may be deployed to
contact with the controller, this may limit large scale sen- reinforce the defenders, so that almost all white communi-
sor deployments. cations are monitored thereafter; but during initial deploy-
In this paper, we will focus on key distribution in com- ment, the white motes have the advantage of surprise.)
modity sensor networks where we do not assume a global It therefore appears of interest to see whether weaken-
passive adversary. Previous work on key distribution for ing our threat model a small proportion of communica-
sensor networks has assumed a strong attacker model: the tions during the network deployment phase and explor-
adversary is assumed to be present both before and after ing whether this can give us useful cost savings and usabil-
node deployment; she can monitor all communications ev- ity gains.
erywhere in the deployment site at all times. It is usual to Using our more realistic attacker model, we design a
assume also that she can subvert (maliciously reprogram) a lightweight security protocol suitable for use in non-critical
small number of targeted sensor nodes. commodity sensor networks. Our protocol is called Key In-
Such assumptions may be appropriate for strate- fection and is based on the assumption that, during the net-
gic sensor-network applications such as nuclear test-ban work deployment phase, the attacker can monitor only a
treaty monitoring, where an international agency em- fixed percentage of communication channels. (We pro-
places sensing devices on the territory of a suspect state. vide a detailed justification for this assumption in Sec-
However, they have led to the development of heavy- tion 3.) We show that, using this slightly relaxed attacker
weight security protocols that incur costly overheads of model, it is possible to perform sufficiently secure key dis-
computation, of storage, or of time and effort needed tribution with low computation overhead (a few symmetric
for pre-deployment configuration. This overhead is al- cryptographic operations), no memory overhead (only stor-
most certainly excessive for more mundane applications. age for the actual keys used in node-to-node communica-
In fact, for many applications, if `security' increases de- tions), and no prior key setup. Due to the lightweight nature
ployment costs or impairs ease of use, then many users of the Key Infection protocol, it is highly suited for imple-
will simply switch it off. For security to be widely used, mentation in low-cost commodity sensor nodes.
it should ship as a default and require no significant ef- This paper makes several contributions. First, we iden-
fort by even non-technical users to configure. For example, tify a more realistic attacker model that is applicable to non-
many early WiFi base stations shipped with encryp- critical commodity sensor networks. Second, we present
tion switched off by default, and many users simply did not Key Infection, a light-weight key-distribution mechanism
turn it on. that is so efficient that it is applicable even to smart dust
We are therefore interested in developing usable secu- sensor nodes. Third, we analyze the security of key infec-
rity, and, when exploring the available trade-offs, one pos- tion, and design Secrecy Amplification, an additional mech-
sible target is the assumption of a global passive adversary. anism to strengthen the security of key infection in the pres-
ence of an active attacker. Together, these provide interest- Another is for the first generation of base stations to pos-
ing new ways to trade off security for cost and usability. Fi- sess master keys that are destroyed once a network has been
nally, the research community has so far largely considered established and link keys have been set up between neigh-
authentication to be a bootstrapping problem. However, in bouring nodes. However, there are applications where this
many real-world applications, the major cost is in mainte- may be unsatisfactory. The deletion of all master keys ef-
nance, rather than initial deployment. We hope this paper fectively closes the network and makes the addition of new
will start to shift the focus toward designing authentication nodes impossible. This makes it difficult to expand a sensor
mechanisms that are optimised over the total lifecycle of a network or to add new nodes to replace failed or battery-
network. exhausted nodes. Also, a number of nodes fail when they
hit the ground, and then be probed by the attacker. So key
erasure cannot be assured.
2. Previous Work
An alternative method has been proposed by Eschenauer
A sensor network is typically set up as follows. When and Gligor [6], and extended by Chan, Perrig, and Song [3].
a node hits the ground, it broadcasts its identity, say i. If a In these key-distribution schemes, enough symmetric keys
neighbouring node j hears it, it replies. The two mutually- are pre-loaded on each node that any two nodes will prob-
aware nodes now set RF power at just the level needed for ably share a key after deployment. However, these schemes
communication. A source-based routing protocol, based on require a significant pre-computation phase in which the
the periodic broadcast of beacons by base stations, orga- shared keys are generated based on the total number of
nizes the nodes into forests, with a base station at the root nodes in the network and the expected density of deploy-
of each tree. This involves not just routing but time syn- ment. There may therefore be significant usability issues.
chronization: to save power, the nodes typically turn off Furthermore, each node requires a lot of memory to store
their communications, only waking up and listening for ra- keys, most of which is effectively wasted since only a small
dio signals intermittently. The base station may be a normal fraction of keys are ever actually used. So there are cost is-
node, or it may have a larger battery so it can cope with a sues too. It appears that random key pre-distribution is not
larger volume of traffic and also communicate with the out- a practical solution for low-cost commodity networks. Re-
side world by transmitting at a higher power. cently, Du et al. [5], and Liu and Ning [11, 12] leverage
Blom's [1] and the Blundo et al. [2] key establishment pro-
The routing architecture depends on the nodes behaving
tocols to achieve high resilience to node capture. Zhu et
themselves that is, executing the software with which they
al. [19] assume that an attacker arrives after key establish-
were loaded before deployment and on the integrity of the
ment, and that all nodes in the network share a secret key.
link traffic between them. If an opponent can intercept and
However, all these protocols assume secret information is
modify this traffic, then he can create disruptions; a typical
set up before sensor network deployment. In this paper we
attack will target the routing mechanism so as to introduce
explore key setup without any prior information.
loops or partition the network. It is also possible to disrupt
the network by maliciously introducing clock skew. Previ-
ous work therefore included services such as authenticated 3. A Real World Attacker Model
broadcast. These rely on shared symmetric keys. The initial
keys are diversified from master keys, known to the base In prior work, researchers have assumed highly capable
stations: thus we may have ki j = {i, j}KM . The base sta- and motivated attacker a global passive adversary, who
tions act as key-distribution centers in much the same way monitors and stores all communications. It is also com-
as with Kerberos, but using protocols designed to minimize monly assumed that the attacker can subvert a small cho-
communications overhead [14]. sen subset of the sensor nodes, and deploy hostile nodes of
While adequate for many environments, this security ar- her own. She is sometimes assumed to be a global active ad-
chitecture is vulnerable to node destruction and subversion. versary, in that she can modify and inject communications
As the base stations send significantly more radio traffic at will at any time.
than the other nodes, an opponent can use direction find- This assumption is inherited from the world of crypto re-
ing to locate them, then either destroy them or subvert them search, which in turn was conditioned by the experience
(for example, by probing them to extract cryptographic keys of World War 2 where both strategic and tactical com-
or to load maliciously altered software). The compromise munications were mostly carried by radio and were rou-
of master keys betrays all traffic that was protected by them tinely intercepted; and the world of international telephony
and that was also overheard and stored by the opponent. in the post-war years where the traffic volumes were such
There are various possible countermeasures. In some ap- that widespread interception by signals intelligence agen-
plications, one can use normal nodes as base stations and cies was feasible and was indeed carried out. However, in
have other nodes replace them after random periods of time. the modern wired world, it is less realistic because of the
cost of tapping high-speed backbones and because of to- have very different attack models, and this is reflected in
day's huge traffic volumes. Intelligence agencies have had their construction each should provide an optimal trade-
to sponsor legislation in many countries to install surveil- off of security and cost for its owner. Homeowners under-
lance devices to get access to traffic near its end-points. In stand that normal doors are not invulnerable to a determined
many countries, the law not only compels communications adversary, but prefer to save money (and enjoy more elegant
service providers (such as phone companies and ISPs) to and convenient access to their homes) by assuming that no-
provide such facilities, but regulates the proportion of sub- body will attack their front door with an acetylene torch or a
scribers which the provider must be able to wiretap simulta- rocket-propelled grenade. Likewise, for commodity sensor
neously. Thus, even if a new application (such as a peer-to- networks, the attacker model should reflect realistic protec-
peer network) is being deployed on a global scale on home tion requirements.
PCs using unencrypted TCP/IP communications, it may of- Our proposed trade-off depends on only a slight weak-
ten be a realistic assumption that no single potential oppo- ening of the attack model that hostile surveillance is not
nent has had the capability to record more than (say) 1% ubiquitous during the deployment phase of the network.
of the initialisation traffic. Of course, if the network is later This phase, the time while the nodes are doing key ex-
deemed to be subversive, its users may be targeted inten- change, may last only several seconds a very small frac-
sively. However, during an initial innocuous deployment, it tion of the network's lifetime. From the attacker's view-
is reasonable to assume that most of the traffic escapes un- point, it would be extremely expensive to deploy surveil-
monitored. lance devices against a large number of offices, factories
Consider now the case mentioned above, of a tactical de- and other targets, and retain them in place for a period of
ployment of 10,000 smart dust motes air-dropped into en- perhaps many years, in order to capture key material. This
emy territory. Assume an initialisation process as the shell would be especially difficult as the main obstacle to surveil-
is armed, whereby a master key KM is generated and trans- lance is the availability of power. Long-term surveillance
mitted to all the motes. On landing, mote i find that mote in general requires either connection to mains electricity, or
j is a neighbour, and uses KM to generate a session key periodic redeployment of battery-operated devices.
{i, j}KM , which they subsequently use. However, some of Thus it is unlikely to be economical for anyone to at-
the motes are broken on impact, and since it is not eco- tack commodity sensor networks by means of universal
nomic to make them tamper-proof, the enemy can probe surveillance directed against premises in general. (Targeted
out KM. The enemy thus has access to all the initialisation surveillance does get attempted against high-value targets
traffic that it recorded. Is this a disaster? Probably not, as such as defence ministries and the residences of heads of
we shall examine in detail below. Unless the defender was government. In case their usual countersurveillance mea-
expecting the motes to be deployed at that location and sures fail, owners of such premises are advised to use wired
therefore had her own defensive motes ready and waiting networks, or to invest in careful cryptographic initialisation
to record communications or insert themselves into the sen- of each node. But theirs is a tiny market and their problems
sor network it would have been impractical for her to have are not our subject here.)
recorded more than a tiny fraction of the initialisation traf- The attacker model we assume is therefore as follows:
fic. So in this case the master key did not achieve anything 1. The attacker does not have physical access to the de-
beyond securing the secrecy of the small fraction of com- ployment site during the deployment phase;
munications that was actually recorded by the adversary. If
this amount of communication information was too little to 2. The attacker is able to monitor only a small proportion
be of significant use to her, then we could have got the same () of the communications of the sensor network dur-
result at lower cost by not using a master key at all, but sim- ing the deployment phase. After key exchange is com-
ply exchanging session keys in the clear. plete, she is able to monitor all communications at will;
Consider now non-critical commodity sensor networks, 3. The attacker is unable to execute active attacks (such
which for cost reasons have extreme limitations on sensor as jamming or flooding) during the deployment phase.
hardware and also require that the pre-deployment setup After key exchange is complete, she is free to launch
must be minimal. These have a quite different threat model any kind of attack.
than tactical military networks, simply because they are less In summary, the attacker is assumed to be fully capable
valuable as targets and little damage is done to the user if at all times except during the deployment phase, where she
their security should fail. Hence, it is rather dubious to ap- is assumed to have at most a partial, passive presence. This
ply a stronger attack model to non-critical commodity sen- is realistic because deployment represents a very small win-
sor networks than we would apply to a tactical military de- dow of opportunity for an adversary. The possibility of an
ployment. attack during the vulnerable window is usually an accept-
As an analogy, doors to bank vaults and doors to homes able risk since the window is extremely small (several sec-
onds) compared to the overall lifetime of the network (up to Assume that node i's signal is heard by node j. Its re-
several years). sponse will be to generate a pairwise key k j and send it,
In order to violate the assumed attacker model, an ad- along with its name, to i: { j, k ji }ki . As we try at all times
versary has to achieve several things. First, she has to have to minimize the energy cost of our protocols, this packet
the foresight to deploy surveillance equipment or adversar- will be transmitted using the minimum power necessary for
ial nodes at the target site before the sensor network is de- the link, based on measurement of the strength of the sig-
ployed there. Second, her eavesdropping devices must re- nal from i.
main in place, operational and undetected, until the sen- The key k ji can be used to protect traffic between i and
sors perform key exchange. Third, she needs to be able to j. It may seem extremely counterintuitive that plaintext key
identify, retrieve and process the the relevant eavesdropped broadcast can give any protection at all. However, in an area
product in order to extract the key exchange messages. In with no opponents, plaintext key exchange foils any adver-
general, except for high-value strategic targets, these re- sary who arrives later. Even where there are opponents al-
quirements will make it too expensive to maintain antici- ready present at the time of deployment, it will still give sig-
patory surveillance against which such targets are vigor- nificant protection. For example, we show below that where
ously defended in any case. there is one `black' (hostile) smart dust sensor node for ev-
When we come to non-critical commodity sensors such ery 100 white nodes, and each node has on average four
as light and temperature sensors for homes, anticipatory neighbours within range, only 2.4% of links will be com-
surveillance makes no sense. If an attacker wishes to vio- promised.
late the privacy of a home by deploying sensors in it, she The situation can be improved considerably by a small
will presumably harvest audio directly, instead of perform- change in the protocol. Instead of each white node broad-
ing sophisticated electromagnetic eavesdropping to steal the casting a single initial key as loudly as it can, it starts off
keys of sensor nodes that just conceivably might be de- transmitting very quietly and steadily increases the power
ployed some time in the unspecified future. until a response is heard. A link key is established with the
Furthermore, since the deployment time window is responder, and then the broadcast resumes with a new initial
small, additional security measures could be taken by net- key. This `key whispering' protocol ensures that two white
work owners whose threat model lies somewhere between nodes W1 and W2 within range of each other and of a black
the random householder's and the Head of Government's. node B will exchange a secure key provided that the black
For example, the site could be secured physically and mon- node is further away from either W1 or W2 than the distance
itored against adversarial entry; it could be swept for between W1 and W2 . In this case, the number of links that
eavesdropping or adversarial devices prior to key ex- the black node can eavesdrop falls to 0.8%. We will now
change; or an intrusion detection system could be run after describe the simulation in more detail.
key exchange to ensure that no external nodes have man-
aged to insert themselves into the network. These additional
measures are simple to execute and relate well to users' ex- 5. Analysis
isting intuitions of security.
Our attack model is in fact probably still too paranoid. Key infection is trivially secure if the attacker arrives af-
However, by using this slightly weakened model, we can ter the key infection phase. In this section, we consider the
greatly simplify key exchange, as we shall describe in the case where an attacker already has some black dust nodes
following section. installed before our customer installs the white dust nodes.
We compute an upper bound on the ratio of communica-
4. Key Infection tion links that the black dust nodes may compromise. As-
sume the maximum range of the radio is R. Let the smart
In view of our real-world attacker model, let us consider dust nodes be distributed over an area of size s, let Nb be
the simplest possible approach to key establishment: each the number of black dust nodes, and let Nw be the number
node simply chooses a key and broadcasts it in plaintext to white dust nodes. Assume the black nodes are distributed
its neighbours. Thus, node i, when it comes to rest, broad- uniformly at random in the area of size s.
casts a key ki . Recall that this is a short-range transmission, For now, we only consider the case where the black
with a maximum range of perhaps ten meters, and perhaps nodes collude. So in order to eavesdrop a key setup between
half a dozen other nodes will have landed within range. As two white nodes, the attacker needs at least one black node
they become active, they detect each others' presence and in the radius of each one of the two white nodes. (In the
start organizing themselves into a network. The idea is to non-colluding case, we do not consider the sharing of infor-
propagate key material as contact is made, rather like an in- mation among black nodes, so the outcome is always more
fection spreading through a biological population. favourable for white.)
If both nodes transmit at maximum strength, then given 6.1. Secrecy amplification
a link e, the effective eavesdropping area is at most R2 , and
hence the probability that this link is bad, i.e., can be eaves- The first strategy is `secrecy amplification' in which we
dropped by at least a black node, is at most R2 Nb /s. combine keys propagated along different paths. Suppose
For the whispering case, if a link has length r, then both that the nodes W1 , W2 , and W3 are neighbours. W1 and W2
nodes will transmit their signals at strength that exactly set up the key k12 , W1 and W3 the key k13 , W2 and W3 the
reaches distance r. Therefore, a black node has to lie in the key k23 . To amplify the secrecy of key k12 , W1 can ask W3
intersection of the two circles of radius r where the distance to exchange an additional key with W2 (here N1 is a unpre-
between the two centers of the two circles is r. The effec- dictable nonce generated by W1 , N2 is a unique nonce gen-
tive eavesdropping area is thus at most the area of this inter- erated by W2 (used for confirmation of key k12 ), and we
. assume that principals can distinguish names, nonces and
section which is 2r 2 ( - 43 ) = 1.2r2 , and the probability
3 keys).
that the link is compromised is at most 1.2r 2 Nb /s.
To evaluate our protocols, we simulated the key infection W1 W3 : {W1 ,W2 , N1 }k13
of a random and uniformly distributed white dust, contam- W3 W2 : {W1 ,W2 , N1 }k23
inated with 1%, 2% or 3% of black dust, and averaged the W2 computes : k12 = H(k12 || N1 )
results over 100 simulations. We simulated 10, 000 white W2 W1 : {N1 , N2 }k12
dust nodes, with 100 eavesdropping black sensor nodes. We W1 W2 : {N2 }k12
assumed that both the white dust node and the black dust
node have a receiver range of 10 meters. We considered var- After this protocol terminates, W1 and W2 update their
ious node densities d, which we characterize by the average key k12 by hashing it with the value just received: k12 =
number of neighbour nodes. H(k12 || N1 ). If k12 was secure before the protocol, the new
We first compared the point-to-point key exchange and k12 will also be secure afterwards. But if the initial link key
whisper mode extension. Table 1 lists the percentage of k12 was compromised, the new one k12 will not be, so long
compromised links for the basic neighbour infection pro- as neither k13 nor k23 is. The last two messages of the pro-
tocol, and the whisper mode extension. We found that tocol are needed for key confirmation, to ensure to W1 and
the whisper mode extension results in approximately three W2 that the other party correctly received the key.
times fewer compromised links.
Tables 2 and 3 show the results of our experiments. We
In the above, we assumed that the black nodes have the simulate the regular neighbour key infection and compare
same receiver sensitivity as the white nodes, which appears it to the secrecy amplification. The tables list the ratio of
reasonable given the economies of scale of single-chip re- compromised links for a varying density of black dust:
ceiver technology. It follows that they would have larger 1%, 2%, and 3%. Table 3 also uses the whispering mode for
batteries or a wired network so they can transmit far- neighbour key infection.
ther. This seems a reasonable strategy for a pre-emplaced So naive three-party secrecy amplification gives an im-
defensive network. However, what the above simulation in- provement of about 20%. Is this worthwhile? Well, it is of-
dicates is that the combatant who can produce the smaller, ten almost free. The reason for this is that nodes which use
denser dust has a significant advantage. optical communications tiny lasers and MEMS corner re-
flectors have mostly unidirectional links, and so need rout-
ing algorithms for finding a loop back to a transmitting node
in any case.
6. Multihop and Multipath Key Establish-
Secrecy amplification is not limited to paths of two hops.
ment For example, to protect against black dust B between the
two white dust nodes W1 and W2 , one can recruit a chain
As the previous section shows, an attacker who intro- of white nodes W3 , W4 , W5 that are out of range of B. The
duces black dust nodes before we deploy our white dust source routing algorithms used in many sensor networks
nodes can subvert some fraction of communication links. give partial location information, and can probably be im-
In this section, we design and analyze Secrecy Amplifica- proved by using link transmitter power as a distance metric.
tion, a technique that utilizes multipath key establishment They can also be easily adapted to search for `a path from A
to make her job significantly harder. We also simulate var- to B that does not pass through C, D or E'. Thus, even when
ious strategies for key establishment in which nodes intro- B is directly between W1 and W2 , there is some hope of find-
duce each other, in an attempt to bypass nodes whose links ing a chain W3 , W4 , W5 that enables W1 to set up a secure key
are compromised from the start (by being too close to black with W2 . The next table has simulations of secrecy amplifi-
nodes) or which are subverted later. cation undertaken using a multihop return path. This is sig-
= 1% = 2% = 3%
d basic whisper basic whisper basic whisper
2 1.13% 0.40% 2.34% 0.81% 3.48% 1.19%
3 1.75% 0.61% 3.51% 1.25% 5.06% 1.81%
4 2.38% 0.83% 4.61% 1.61% 6.75% 2.44%
5 2.92% 1.01% 5.76% 2.00% 8.40% 3.02%
Table 1: This table compares the standard key infection with the whisper-mode key infection. The first column lists d, the
average number of neighbours of a node. The remaining columns list the ratio of compromised links (links that a black dust
mote could eavesdrop on and extract both key infection messages).
= 1% = 2% = 3%
d basic SA basic SA basic SA
2 1.20% 0.97% 2.29% 2.00% 3.38% 2.93%
3 1.81% 1.37% 3.44% 2.67% 5.42% 3.93%
4 2.30% 1.80% 4.45% 3.71% 6.50% 5.55%
5 2.93% 2.37% 5.73% 4.68% 8.73% 6.75%
Table 2: This table shows the improvement of secrecy amplification (SA) over the basic key infection.
nificantly better where complexity and other constraints above; and as for routing, a number of multihop keying
permit it. strategies are available, which require nodes to store more
keys than they have neighbours but very much fewer than
there are nodes.
6.2. Multihop keys
6.3. Interaction with routing algorithms
The second strategy is setting up multihop keys. If W1
links to W2 which links to W3 , then it may make sense for Some existing work on secure ad-hoc routing assumes
W1 and W3 to invoke W2 's help to set up a key that W2 im- a particular routing strategy. Our work does not; existing
mediately forgets, against the eventuality that W2 is com- prototypes use strategies based on dynamic source rout-
promised in the future. This has two purposes. First, it sup- ing [13, 8] although our key infection protocol can also sup-
ports end-to-end, rather than link-level, cryptography. It is port other mechanisms.
energy efficient for base-to-node communications to be en- These other mechanisms may have to be used to recover
crypted using end-to-end keys rather than translated at in- from attacks. For example, if sufficient nodes are subverted
tervening nodes; this also protects against subsequent node for the network to be partitioned that is, there are pairs
compromise. Second, multihop keying also protects multi- of motes that can no longer route to each other despite be-
hop secrecy amplification against node compromise. ing physically connected by a multihop path then a recov-
Where memory is not too restricted, multihop keying ery phase may be initiated. This may involve backup nodes,
may be a very natural mechanism to use. Consider the un- if they exist; a re-run of the initial network discovery algo-
solved problem mentioned by Sirois and Kent [16], that se- rithm; or a strategy such as sticky random routing. There,
cure routing protocols that only use link-level protection we use a single path until it becomes unavailable, whether
(such as Nimrod with IPSEC) are vulnerable to node sub- as a result of congestion or (in our application) damage,
version. The point is that, in a `proper' network, a router and then switch to an alternate [10]. Our multipath key in-
has the memory to set up a key with every other router for fection protocol automatically discovers paths that may be
which it keeps a routing table entry. used for this as needed. (This is one point at which the anal-
In sensor networks such as Smart Dust, memory size and ogy with biological infection breaks down. In biology, the
the energy cost of messages are the critical resource lim- immune response normally stops you catching the same dis-
its. However, sensor networks have quite limited types of ease twice; if you are a smart dust mote, the more keys you
traffic mostly application messages between base stations `catch' from a colleague, the better.)
and nodes, local routing / key management messages, and In addition, multihop keying enables motes to try differ-
broadcasts of signals such as time beacons. Base-to-node ent logical paths along the same physical path, in order to
traffic should be end-to-end encrypted anyway, as noted identify and isolate a faulty or subverted node. If base sta-
= 1% = 2% = 3%
d basic SA basic SA basic SA
2 0.38% 0.34% 0.75% 0.66% 1.25% 1.08%
3 0.59% 0.49% 1.15% 0.93% 1.75% 1.50%
4 0.81% 0.61% 1.67% 1.33% 2.27% 1.87%
5 1.04% 0.81% 2.03% 1.53% 3.15% 2.28%
Table 3: This table shows the improvement of secrecy amplification (SA) over the basic key infection. In this case, the basic
key infection uses whispering.
= 1% = 2% = 3%
d basic m-path basic m-path basic m-path
2 0.61% 0.38% 1.40% 0.48% 2.23% 1.11%
3 0.55% 0.26% 1.11% 0.58% 1.76% 0.91%
4 0.40% 0.16% 0.94% 0.30% 1.57% 0.80%
5 0.35% 0.04% 0.75% 0.12% 1.29% 0.40%
Table 4: This table compares the basic two-hop key infection, with the multipath extension. We simulate nodes that perform
key infection with neighbours that are two hops away. In the columns marked "basic", we assume that the return path of the
key infection is the same as the forward path. In the columns marked "m-path", the return message takes a different path, if
available.
tion W1 communicates with a sensor W5 via the path W2 , W3 , out a key, so a network-wide bootstrap key will be used.
W4 , then normally routing updates would be broadcast using However, our discussion may perhaps have convinced the
a key shared with all these nodes [14]. However, more se- reader that this is only a small part of the equation. (Even
lective multicasts can be tried as a first recourse in the event if initial keying is used, and if the motes are moderately
of failure. tamper-resistant, the use of the mechanisms described here
Finally, although most current sensor networks do not may push down the value of recovering the bootstrap key
need to do mobile routing, the topology does still change below the cost of probing it out.)
as a result of battery exhaustion and node destruction. In Given an opponent with the capability to subvert nodes
the future, we will need routing strategies that work for after deployment, the balance between attack and defense
much more mobile principals, such as swarms of insect will move towards an equilibrium that will depend on the
robots (which might be flying around gobbling up hostile cost-benefit ratios of attack and of security maintenance. If
dust motes). these favor the defender, the attacker will give up; otherwise
there will be an equilibrium at which the defender has to go
7. Economic Issues all out [7, 18].
One implication is that, if the deployment is to be at
When analyzing the economics of a game, one approach all long lived, the initial costs for both parties will be
is to look at the initial and marginal costs borne by each dwarfed by the running costs. The designer of the white
party. Thus, a company using a larger-scale production pro- dust should focus on the cost of security maintenance rather
cess than a competitor may have higher initial costs but than on over-investing in the initial trust bootstrap mecha-
lower marginal costs. Thus it may be dangerous for a small nisms. This mistake, though common enough, is from the
company to develop a market to the point at which it be- viewpoint of an economic model somewhat like placing too
comes attractive to a large competitor. much trust in security-by-obscurity. Security maintenance
A similar effect appears here. If initial keying is used, and renewability is often more important, in long-lived sys-
this imposes a cost on the party deploying the dust (e.g., tems, than the height of the initial barrier placed before in-
having a mechanism to distribute a seed key to the nodes truders.
in a canister just before deployment). It also imposes a cost
on the attacker; in addition to having to pre-emplace black 8. Conclusions
nodes with a certain density, he also has to physically re-
cover a seed key from a white node. In practice, it may well We proposed a novel and quite counterintuitive way of
be cheaper for white to key her nodes than for black to probe managing keys in sensor networks: each node bootstraps it-
self by broadcasting an initial key in the clear. Nodes then more similar to societal mechanisms than to existing indus-
exchange keys and build up trust structures as they do net- trial control technology. The command and control of sen-
work and resource discovery. It turns out that, under often sor networks appears to be a first step along the way to de-
reasonable assumptions, this is almost as secure as using veloping them.
pre-loaded initial keys. In other words, initial keying in such
networks buys less than one might think. Acknowledgment:
Up till now, the research community has focused on
mechanisms for establishing initial keys. This paper has This work was done while the first author was visit-
shown how the benefits of initial keying can be analyzed ing UC Berkeley on sabbatical in 2001-2, supported by
separately from the benefits of later-stage key management NSF 9979852. The other authors are supported by NSF
activities, such as key updating, the use of alternative trust CNS-0347807 (CAREER) and by a gift from Bosch.
routes, and the invocation of backups whether in response
to perceived attacks, or periodically. Such resilience and re- References
covery mechanisms are often more important than boot-
strapping, but are widely ignored as operational matters [1] R. Blom. Non-public key distribution. In Advances in Cryp-
about which there is little of academic interest to be said. tology: Proceedings of Crypto '82, pages 231236, 1982.
We hope that contemplating systems without an initial trust- [2] C. Blundo, A. D. Santis, A. Herzberg, S. Kutten, U. Vaccaro,
worthy keying phase may be a good way to challenge this and M. Yung. Perfectly-secure key distribution for dynamic
mindset. conferences. In Advances in Cryptology - Crypto '92, pages
471486, 1992.
Furthermore, the protection of sensor networks is a suf-
[3] H. Chan, A. Perrig, and D. Song. Random key predistribu-
ficiently compact problem to be accessible to network mod- tion schemes for sensor networks. In IEEE Symposium on
elling techniques. We can explore, in a quantitative way, the Security and Privacy, May 2003.
benefits of differing schemes for initial keying, key updat- [4] W. Diffie and M. Hellman. New directions in cryptography.
ing and recovery from failure. Key infection is thus of in- IEEE Transactions on Information Theory, IT-22(6):644
terest not just as a curiosity, but as a tool for understanding 654, Nov. 1976.
more complex systems. [5] W. Du, J. Deng, Y. Han, and P. Varshney. A pairwise key pre-
There are many environments in which an opponent can distribution scheme for wireless sensor networks. In Pro-
compromise any principal, but cannot compromise all of ceedings of the Tenth ACM Conference on Computer and
them. For example, a peer-to-peer system may be modeled Communications Security (CCS 2003), pages 4251, Oct.
as a large self-organizing network of principals, any small 2003.
proportion of which can be subverted at will by the oppo- [6] L. Eschenauer and V. D. Gligor. A key-management scheme
for distributed sensor networks. In Proceedings of the 9th
nent. As such systems attract hostile action once they have
ACM Conference on Computer and Communication Secu-
grown to a certain size, the bootstrapping issues are not so
rity, pages 4147, Nov. 2002.
important as what happens later.
[7] J. Hirshleifer. Economic Behaviour in Adversity. Chicago
What cryptography mainly achieves in such systems is UP, 1987.
to stop security compromises becoming worse over time. [8] Y.-C. Hu, A. Perrig, and D. B. Johnson. Ariadne: A se-
Whatever bad nodes managed to join the system at its in- cure on-demand routing protocol for ad hoc networks. In
ception remain there, until they show themselves (and hope- Proceedings of the Eighth ACM International Conference on
fully get removed by higher layer mechanisms). If they Mobile Computing and Networking (Mobicom 2002), Sept.
choose not to show themselves, then in some applications 2002.
(such as routing, where it's integrity that matters) they can [9] J. M. Kahn, R. H. Katz, and K. S. Pister. Mobile networking
perhaps be ignored. for smart dust. In International Conference on Mobile Com-
Analogies with biological systems and even with human puting and Networking (MobiCom '99), Seattle, WA, August
1999.
social structures may also be useful. The trust propagation
[10] F. P. Kelly. Modeling communication networks, present and
and maintenance mechanisms described here mirror those
future. Philosophical Transactions of the Royal Society,
in human society. You can almost certainly not remember A354:437463, 1996.
the time when you decided to trust your mother! Initial trust [11] D. Liu and P. Ning. Establishing pairwise keys in distributed
bootstrapping is not such a big issue in many human organ- sensor networks. In Proceedings of the Tenth ACM Con-
isations as stiffening members against later subversion. If ference on Computer and Communications Security (CCS
our future is to be built on huge ad-hoc networks of commu- 2003), pages 5261, Oct. 2003.
nicating smart objects, from swarms of robot insects down [12] D. Liu and P. Ning. Location-based pairwise key establish-
to nanites circulating in our bloodstream, then the mecha- ments for static sensor networks. In ACM Workshop on Secu-
nisms we use to command and control them may be much rity in Ad Hoc and Sensor Networks (SASN '03), Oct. 2003.
[13] R. Perlman. Network Layer Protocol with Byzantine Agree-
ment. PhD thesis, The MIT Press, Oct. 1988. LCS TR-429.
[14] A. Perrig, R. Szewczyk, V. Wen, D. Culler, and J. D. Tygar.
SPINS: Security protocols for sensor networks. In Proceed-
ings of ACM International Conference on Mobile Computing
and Networks (MobiCom 2001), pages 189199, July 2001.
[15] R. Rivest, A. Shamir, and L. Adleman. A method for obtain-
ing digital signatures and public-key cryptosystems. Com-
munications of the ACM, 21(2):120126, Feb. 1978.
[16] K. Sirois and S. Kent. Securing the nimrod routing architec-
ture. In Proceedings of the Symposium on Network and Dis-
tributed Systems Security (NDSS '97). Internet Society, Feb.
1997.
[17] F. Stajano and R. Anderson. The resurrecting duckling:
Security issues for ad-hoc wireless networks. In Secu-
rity Protocols--7th International Workshop, volume 1796
of Lecture Notes in Computer Science, pages 172194.
Springer-Verlag, Berlin Germany, 2000.
[18] H. R. Varian. System reliability and free riding. In Workshop
on Economics and Information Security, May 2002.
[19] S. Zhu, S. Setia, and S. Jajodia. LEAP: Efficient security
mechanisms for large-scale distributed sensor networks. In
Proceedings of the Tenth ACM Conference on Computer and
Communications Security (CCS 2003), pages 6272, Oct.
2003.