MISFIT/PIN Cracking - UCAM-CL-TR-560_pdf.md

# PIN Cracking - UCAM-CL-TR-560


---

Technical Report
UCAM-CL-TR-560
ISSN 1476-2986
Number 560
Computer Laboratory
Decimalisation table attacks for PIN
cracking
Mike Bond, Piotr Zielin·ski
February 2003
15 JJ Thomson Avenue
Cambridge CB3 0FD
United Kingdom
phone +44 1223 763500
http://www.cl.cam.ac.uk/

c 2003 Mike Bond, Piotr Zielin·ski
(cid:13)
Technical reports published by the University of Cambridge
Computer Laboratory are freely available via the Internet:
http://www.cl.cam.ac.uk/TechReports/
Series editor: Markus Kuhn
ISSN 1476-2986

Decimalisation table attacks for PIN cracking
Mike Bond, Piotr Zielin(cid:19)ski
Abstract
We present an attack on hardware security modules used by retail banks for the
secure storage and veri(cid:12)cation of customer PINs in ATM (cash machine) infrastruc-
tures. By using adaptive decimalisation tables and guesses, the maximum amount
of information is learnt about the true PIN upon each guess. It takes an average of
15 guesses to determine a four digit PIN using this technique, instead of the 5000
guesses intended. In a single 30 minute lunch-break, an attacker can thus discover
approximately 7000 PINs rather than 24 with the brute force method. With a $300
withdrawal limitpercard, thepotentialbountyisraisedfrom$7200to$2.1million
and a single motivated attacker could withdraw $30{50 thousand of this each day.
This attack thus presents a serious threat to bank security.
1 Introduction
Automatic Teller Machines (ATMs) are used by millions of customers every day to make
cash withdrawals from their accounts. However, the wide deployment and sometimes
secludedlocationsofATMsmakethemidealtoolsforcriminalstoturntraceableelectronic
money into clean cash.
The customer PIN is the primary security measure against fraud; forgery of the mag-
netic stripe on cards is trivial in comparison to PIN acquisition. A street criminal can
easily steal a cash card, but unless he observes the customer enter the PIN at an ATM,
he can only have three guesses to match against a possible 10,000 PINs and would rarely
strike it lucky. Even when successful, his theft still cannot exceed the daily withdrawal
limit of around $300 . However, bank programmers have access to the computer systems
tasked with the secure storage of PINs, which normally consist of a mainframe connected
to a \Hardware Security Module" (HSM) which is tamper-resistant and has a restricted
API such that it will only respond to with a YES/NO answer to a customer’s guess.
A crude method of attack is for a corrupt bank programmer to write a program that
tries all PINs for a particular account, and with average luck this would require about
5000 transactions to discover each PIN. A typical HSM can check maybe 60 trial PINs
per second in addition to its normal load, thus a corrupt employee executing the program
during a 30 minute lunch break could only make o(cid:11) with about 25 PINs.
However, HSMs implementing several common PIN generation methods have a (cid:13)aw.
The (cid:12)rst ATMs were IBM 3624s, introduced widely in the US in around 1980, and most
PIN generation methods are based upon their approach. They calculate the customer’s
original PIN by encrypting the account number printed on the front of the customer’s
card with a secret DES key called a \PIN generation key". The resulting ciphertext
3

is converted into hexadecimal, and the (cid:12)rst four digits taken. Each digit has a range of
‘0’-‘F’.InordertoconvertthisvalueintoaPINwhichcanbetypedonadecimalkeypad,
a \decimalisation table" is used, which is a many-to-one mapping between hexadecimal
digits and numeric digits. The left decimalisation table in Figure 1 is typical.
0123456789ABCDEF 0123456789ABCDEF
0123456789012345 0000000100000000
Figure 1: Normal and attack decimalisation tables
This table is not considered a sensitive input by many HSMs, so an arbitrary table
can be provided along with the account number and a trial PIN. But by manipulating
the contents of the table it becomes possible to learn much more about the value of the
PIN than simply excluding a single combination. For example, if the right hand table is
used, a match with a trial pin of 0000 will con(cid:12)rm that the PIN does not contain the
number 7, thus eliminating over 10% of the possible combinations. We (cid:12)rst present a
simple scheme that can derive most PINs in around 24 guesses, and then an adaptive
scheme which maximises the amount of information learned from each guess, and takes
an average of 15 guesses. Finally, a third scheme is presented which demonstrates that
the attack is still viable even when the attacker cannot control the guess against which
the PIN is matched.
Section 2 of the paper sets the attack in the context of a retail banking environment,
and explains why it may not be spotted by typical security measures. Section 3 de-
scribes PIN generation and veri(cid:12)cation methods, and section 4 describes the algorithms
we have designed in detail. We present our results from genuine trials in section 5, discuss
preventative measures in section 6, and draw our conclusions in section 7.
2 Banking Security
Banks have traditionally led the way in (cid:12)ghting fraud from both insiders and outsiders.
They have developed protection methods against insider fraud including double-entry
book-keeping, functional separation, and compulsory holiday periods for sta(cid:11), and they
recognise the need for regular security audits. These methods successfully reduce fraud
to an acceptable level for banks, and in conjunction with an appropriate legal framework
for liability, they can also protect customers against the consequences of fraud.
However, the increasing complexity of bank computer systems has not been accom-
panied by su(cid:14)cient development in understanding of fraud prevention methods. The
introduction of HSMs to protect customer PINs was a step in the right direction, but
even in 2002 these devices have not been universally adopted, and those that are used
have been shown time and time again not to be impervious to attack [1, 2, 5]. Typical
banking practice seeks only to reduce fraud to an acceptable level, but this translates
poorly into security requirements; it is impossible to accurately assess the security expo-
sure of a given (cid:13)aw, which could be an isolated incident or the tip of a huge iceberg. This
sort of risk management con(cid:13)icts directly with modern security design practice where ro-
bustness is crucial. There are useful analogues in the design of cryptographic algorithms.
Designers who make \just-strong-enough" algorithms and trade robustness for speed or
4

export approval play a dangerous game. The cracking of the GSM mobile phone cipher
A5 is but one example [3].
And as \just-strong-enough" cryptographic algorithms continue to be used, the risk
of fraud from brute force PIN guessing is still considered acceptable, as it should take
at least 10 minutes to guess a single PIN at the maximum transaction rate of typical
modules deployed in the 80s. Customers are expected to notice the phantom withdrawals
and report them before the attacker could capture enough PINs to generate a signi(cid:12)cant
liability for the banks. Even with the latest HSMs that support a transaction rate ten
times higher, the sums of money an attacker could steal are small from the perspective
of a bank.
But now that the PIN decimalisation table has been identi(cid:12)ed as an security relevant
data item, and the attacks described in this paper show how to exploit uncontrolled access
to it, brute force guessing is over two orders of magnitude faster. Enough PINs to unlock
access to over $2 million can be stolen in one lunch break!
A more sinister threat is the perpetration of a smaller theft, where the necessary
transactions are well camou(cid:13)aged within the banks audit trails. PIN veri(cid:12)cations are
not necessarily centrally audited at all, and if we assume that they are, the 15 or so
transactions required will be hard for an auditor to spot amongst a stream of millions.
Intrusion detection systems do not fare much better { suppose a bank has an extremely
strict audit system that tracks the number of failed guesses for each account, raising
an alarm if there are three failures in a row. The attacker can discover a PIN without
raising the alarm by inserting the attack transactions just before genuine transactions
from the customer which will reset the count. No matter what the policies of the intrusion
detection system it is impossible to keep them secret, thus a competent programmer could
evade them. The very reason that HSMs were introduced into banks was that mainframe
operating systems only satisfactorily protected data integrity, and could not be trusted
to keep data con(cid:12)dential from programmers.
So as the economics of security (cid:13)aws like these develops into a mature (cid:12)eld, it seems
that banks need to update their risk management strategies to take account of the volatile
nature of the security industry. They also have a responsibility to their customers to
reassess liability for fraud in individual cases, as developments in computer security con-
tinually reshape the landscape over which legal disputes between bank and customer are
fought.
3 PIN Generation & Veri(cid:12)cation Techniques
There are a number of techniques for PIN generation and veri(cid:12)cation, each proprietary
to a particular consortium of banks who commissioned a PIN processing system from a
di(cid:11)erent manufacturer. The IBM CCA supports a representative sample, shown in Figure
2. We IBM 3624-O(cid:11)set method in more detail as it is typical of decimalisation table use.
3.1 The IBM 3624-O(cid:11)set PIN Derivation Method
The IBM 3624-O(cid:11)set method was developed to support the (cid:12)rst generation of ATMs and
has thus been widely adopted and mimicked. The method was designed so that o(cid:15)ine
ATMs would be able to verify customer PINs without needing the processing power and
5

Method Uses Dectables
IBM 3624 yes
IBM 3624-O(cid:11)set yes
Netherlands PIN-1 yes
IBM German Bank Pool Institution yes
VISA PIN-Validation Value
Interbank PIN
Figure 2: Common PIN calculation methods
storage to manipulate an entire database of customer account records. Instead, a scheme
was developed where the customer’s PIN could be calculated from their account number
by encryption with a secret key. The account number was made available on the magnetic
stripe of the card, so the ATM only needed to securely store a single cryptographic key.
An example PIN calculation is shown in Figure 4.
The account number is represented using ASCII digits, and then interpreted as a
hexadecimal input to the DES block cipher. After encryption with the secret \PIN gen-
eration" key, the output is converted to hexadecimal, and all but the (cid:12)rst four digits
are discarded. However, these four digits might contain the hexadecimal digits ‘A’-‘F’,
which are not available on a standard numeric keypad and would be confusing to cus-
tomers, so they are mapped back to decimal digits using a \decimalisation table" (Figure
3).
0123456789ABCDEF
0123456789012345
Figure 3: A typical decimalisation table
Account Number 4556 2385 7753 2239
Encrypted Accno 3F7C 2201 00CA 8AB3
Shortened Enc Accno 3F7C
0123456789ABCDEF
0123456789012345
Decimalised PIN 3572
Public Offset 4344
Final PIN 7816
Figure 4: IBM 3624-O(cid:11)set PIN Generation Method
6

The example PIN of 3F7C thus becomes 3572. Finally, to permit the cardholders to
change their PINs, an o(cid:11)set is added which is stored in the mainframe database along
with the account number. When an ATM veri(cid:12)es an entered PIN, it simply subtracts
the o(cid:11)set from the card before checking the value against the decimalised result of the
encryption.
3.2 Hardware Security Module APIs
Bank control centres and ATMs use Hardware Security Modules (HSMs), which are
charged with protecting PIN derivation keys from corrupt employees and physical at-
tackers. An HSM is a tamper-resistant coprocessor that runs software providing crypto-
graphic and security related services. Its API is designed to protect the con(cid:12)dentiality
and integrity of data while still permitting access according to a con(cid:12)gurable usage policy.
Typical (cid:12)nancial APIs contain transactions to generate and verify PINs, translate guessed
PINsbetweendi(cid:11)erentencryptionkeysastheytravelbetweenbanks, andsupportawhole
host of key management functions.
The usage policy is typically set to allow anyone with access to the host computer
to perform everyday commands such as PIN veri(cid:12)cation, but to ensure that sensitive
functionssuchasloadingnewkeyscanonlybeperformedwithauthorisationfrommultiple
employees who are trusted not to collude.
IBM’s \Common Cryptographic Architecture" [6] is a (cid:12)nancial API implemented by a
range of IBM HSMs, including the 4758, and the CMOS Cryptographic Coprocessor (for
PCs and mainframes respectively). An example of the code for a CCA PIN veri(cid:12)cation
is shown in Figure 5.
Encrypted_PIN_Verify(
A_RETRES , A_ED , // return codes 0,0=yes 4,19=no
trial_pin_kek_in , pinver_key , // encryption keys for enc inputs
(UCHAR*)"3624 " "NONE " // PIN block format
" F" // PIN block pad digit
(UCHAR*)" " ,
trial_pin , // encrypted_PIN_block
I_LONG(2) ,
(UCHAR*)"IBM-PINO" "PADDIGIT" , // PIN verification method
I_LONG(4) , // # of PIN digits = 4
"0123456789012345" // decimalisation table
"123456789012 " // PAN_data (account number)
"0000 " // offset data
);
Figure 5: Sample code for PIN veri(cid:12)cation in CCA
The crucial inputs to Encrypted_PIN_Verify are the decimalisation table, the
PAN_data, and the encrypted_PIN_block. The (cid:12)rst two are supplied in the clear and are
straightforward for the attacker to manipulate, but obtaining an encrypted_PIN_block
that represents a chosen trial PIN is rather harder.
7

3.3 Obtaining chosen encrypted trial PINs
Some bank systems permit clear entry of trial PINs from the host software. For instance,
this functionality may be required to input random PINs when generating PIN blocks
for schemes that do not use decimalisation tables. The appropriate CCA command is
Clear_PIN_Encrypt, which will prepare an encrypted PIN block from the chosen PIN. It
should be noted that enabling this command carries other risks as well as permitting our
attacks. If thereis notrandomisedpaddingof PINs before theyareencrypted, anattacker
could make a table of known trial encrypted PINs, compare each arriving encrypted PIN
against this list, and thus easily determine its value. If it is still necessary to enable clear
PIN entry in the absence of randomised padding, some systems can enforce that the clear
PINs are only encrypted under a key for transit to another bank { in which case the
attacker cannot use these guesses as inputs to the local veri(cid:12)cation command.
So, under the assumption that clear PIN entry is not available to the attacker, his
second option is to enter the required PIN guesses at a genuine ATM, and intercept the
encrypted PIN block corresponding to each guess as it arrives at the bank. Our adaptive
decimalisation table attack only requires (cid:12)ve di(cid:11)erent trial PINs { 0000 , 0001 ,0010 ,
0100 , 1000. However the attacker might only be able to acquire encrypted PINs under
a block format such as ISO-0, where the account number is embedded within the block.
This would require him to manually input the (cid:12)ve trial PINs at an ATM for each account
that could be attacked { a huge undertaking which totally defeats the strategy.
A third and more most robust course of action for the attacker is to make use of the
PIN o(cid:11)set capability to convert a single known PIN into the required guesses. This known
PIN might be discovered by brute force guessing, or simply opening an account at that
bank.
Despite all these options for obtaining encrypted trial PINs it might be argued that
the decimalisation table attack is not exploitable unless it can be performed without a
singleknowntrialPIN.Toaddresstheseconcerns, wecreatedathirdalgorithm(described
in the next section), which is of equivalent speed to the others, and does not require any
known or chosen trial PINs.
4 Decimalisation Table Attacks
In this section, we describe three attacks. First, we present a 2-stage simple static scheme
which needs only about 24 guesses on average. The shortcoming of this method is that
it needs almost twice as many guesses in the worst case. We show how to overcome this
di(cid:14)culty by employing an adaptive approach and reduce the number of necessary guesses
to 22. Finally, we present an algorithm which uses PIN o(cid:11)sets to deduce a PIN from a
single correct encrypted guess, as is typically supplied by the customer from an ATM.
4.1 Initial Scheme
The initial scheme consists of two stages. The (cid:12)rst stage determines which digits are
present in the PIN. The second stage consists in trying all the possible pins composed of
those digits.
8

Let D be the original decimalisation table. For a given digit i, consider a binary
orig
decimalisation table D with the following property. The table D has 1 at position x if
i i
and only if D has the digit i at that position. In other words,
orig
1 if D [x] = i;
orig
D [x] =
i
(0 otherwise:
For example, for a standard table D = 0123456789012345, the value of D is
orig 3
0001000000000100.
In the (cid:12)rst phase, for each digit i, we check the original PIN against the decimalisation
table D with a trial PIN of 0000. It is easy to see that the test fails exactly when the
i
original PIN contains thedigit i. Thus, using only at most 10 guesses, we have determined
all the digits that constitute the original PIN.
In the second stage we try every possible combination of those digits. Their actual
number depends on how many di(cid:11)erent digits the PIN contains. The table below gives
the details.
Digits Possibilities
A AAAA(1)
AB ABBB(4), AABB(6), AAAB(4)
ABC AABC(12), ABBC(12), ABCC(12)
ABCD ABCD(24)
The table shows that the second stage needs at most 36 guesses (when the original
PIN contains 3 di(cid:11)erent digits), which gives 46 guesses in total. The expected number of
guesses is, however, as small as about 23:5.
4.2 Adaptive Scheme
The process of cracking a PIN can be represented by a binary search tree. Each node v
contains a guess, i.e., a decimalisation table D and a pin p . We start at the root node
v v
and go down the tree along the path that is determined by the results of our guesses. Let
p be the original PIN. At each node, we check whether D (p ) = p . Then, we move
orig v orig v
to the right child if yes and to the left child otherwise.
Each node v in the tree can be associated with a list of original PINs such that
v
P
p if and only if v is reached in the process described in the previous paragraph if we
v
2 P
take p as the original PIN. In particular, the list associated with the root node contains
all possible pins and the list of each leaf should contain only one element: an original PIN
p .
orig
Consider the initial scheme described in the previous section as an example. For
simplicityassumethattheoriginalPINconsistsoftwobinarydigitsandthedecimalisation
table is trivial and maps 0 0 and 1 1. Figure 6 depicts the search tree for these
! !
settings.
Themaindrawbackoftheinitialschemeisthatthenumberofrequiredguessesdepends
strongly on the original PIN p . For example, the method needs only 9 guesses for
orig
p = 9999 (because after ascertaining that digit 0{8 do not occur in p this is the only
orig orig
9

D (p)=? 00
10
yes no
p=11 D (p)=? 10
01
yes no
p=10 D (p)=? 01
01
yes no
p=01 p=00
Figure 6: The search tree for the initial scheme. D denotes the decimalisation table
xy
that maps 0 x and 1 y.
! !
possibility), but there are cases where 46 guesses are required. As a result, the search tree
is quite unbalanced and thus not optimal.
One method of producing a perfect search tree (i.e., the tree that requires the smallest
possible numbers of guesses in the worst case) is to consider all possible search trees and
choose the best one. This approach is, however, prohibitively ine(cid:14)cient because of its ex-
ponential time complexity with respect to the number of possible PINs and decimalisation
tables.
It turns out that not much is lost when we replace the exhaustive search with a simple
heuristics. We will choose the values of D and p for each node v in the following manner.
v v
Let be the list associated with node v. Then, we look at all possible pairs of D and
v v
P
p and pick the one for which the probability of D (p) = p for p is as close to 1 as
v v v 2 P v 2
possible. This ensures that the left and right subtrees are approximately of the same size
so the whole tree should be quite balanced.
This scheme can be further improved using the following observation. Recall that the
original PIN p is a 4-digit hexadecimal number. However, we do not need to determine
orig
it exactly; all we need is to learn the value of p = D (p ). For example, we do not
orig orig
need to be able to distinguish between 012D and ABC3 because for both of them p = 0123.
It can be easily shown that we can build the search tree that is based on the value of p
instead of p provided that the tables D do not distinguish between 0 and A, 1 and B
orig v
and so on. In general, we require each D to satisfy the following property: for any pair
v
of hexadecimal digits x, y: D [x] = D [y] must imply D [x] = D [y]. This property
orig orig v v
is not di(cid:14)cult to satisfy and in reward we can reduce the number of possible PINs from
164 = 65536 to 104 = 10000. Figure 7 shows a sample run of the algorithm for the
original PIN p = 3491.
orig
4.3 PIN O(cid:11)set Adaptive Scheme
When the attacker does not know any encrypted trial PINs, and cannot encrypt his own
guesses, he can still succeed by manipulating the o(cid:11)set parameter used to compensate for
customer PIN change. Our (cid:12)nal scheme has the same two stages as the initial scheme, so
10

No Possible pins Decimalisation table D Trial pin p D (p ) p =? D (p )
v v v orig v v orig
1 10000 1000001000100000 0000 0000 yes
2 4096 0100000000010000 0000 1000 no
3 1695 0111110000011111 1111 1011 no
4 1326 0000000100000000 0000 0000 yes
5 736 0000000010000000 0000 0000 yes
6 302 0010000000001000 0000 0000 yes
7 194 0001000000000100 0000 0001 no
8 84 0000110000000011 0000 0010 no
9 48 0000100000000010 0000 0010 no
10 24 0100000000010000 1000 1000 yes
11 6 0001000000000100 0100 0001 no
12 4 0001000000000100 0010 0001 no
13 2 0000100000000010 0100 0010 no
Figure 7: Sample output from adaptive test program
our (cid:12)rst task is to determine the digits present in the PIN.
Assume that an encrypted PIN block containing the correct PIN for the account has
been intercepted (the vast majority of arriving encrypted PIN blocks will satisfy this
criterion), and for simplicity that the account holder has not changed his PIN and the
correct o(cid:11)set is 0000. Using the following set of decimalisation tables, the attacker can
determine which digits are present in the correct PIN.
D [x]+1 if D [x] = i;
orig orig
D [x] =
i
(D
orig
[x] otherwise:
For example, for D = 0123456789012345, the value of D is 0124456789012445.
orig 3
He supplies the correct encrypted PIN block and the correct o(cid:11)set each time.
As with the initial scheme, the second phase determines the positions of the digits
present in the PIN, and is again dependent upon the number of repeated digits in the
originalPIN.ConsiderthecommoncasewhereallthePINdigitsaredi(cid:11)erent, forexample
1583. We can try to determine the position of the single 8 digit by applying an o(cid:11)set to
di(cid:11)erent digits and checking for a match.
Guess Guess Customer Customer Guess Decimalised Verify
O(cid:11)set Decimalisation Table Guess + Guess O(cid:11)set Original PIN Result
0001 0123456799012345 1583 1584 1593 no
0010 0123456799012345 1583 1593 1593 yes
0100 0123456799012345 1583 1683 1593 no
1000 0123456799012345 1583 2583 1593 no
Each di(cid:11)erent guessed o(cid:11)set maps the customer’s correct guess to a new PIN which
may or may not match the original PIN after it is decimalised using the modi(cid:12)ed table.
This procedure is repeated until the position of all digits is known. Cases with all digits
di(cid:11)erent will require at most 6 transactions to determine all the position data. Three
11

di(cid:11)erent digits will need a maximum of 9 trials, two digits di(cid:11)erent up to 13 trials, and
if all the digits are the same no trials are required as there are no permutations. When
the parts of the scheme are assembled, 16.5 guesses are required on average to determine
a given PIN.
5 Results
We (cid:12)rst tested the adaptive algorithm exhaustively on all possible PINs. The distribution
in Figure 8 was obtained. The worst case has been reduced from 45 guesses to 24 guesses,
and the average has fallen from 24 to 15 guesses. We then implemented the attacks
on the IBM Common Cryptographic Architecture (version 2.41, for the IBM 4758), and
successfully extracted PINs generated using the IBM 3624 method. We also checked the
attacks against the API speci(cid:12)cations for the VISA Security Module (VSM) , and found
them to be e(cid:11)ective. The VSM is the forerunner of a whole range of hardware security
modules for PIN processing, and we believe that the attacks will also be e(cid:11)ective against
many of its successors.
2500
2000
1500
1000
500
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Number of Attempts
sNIP
fo
rebmuN
Figure 8: Distribution of guesses required using adaptive algorithm
12

6 Prevention
It is easy to perform a check upon the validity of the decimalisation table. Several
PIN veri(cid:12)cation methods that use decimalisation tables require that the table should
be 0123456789012345 for the algorithm to function correctly, and in these cases the API
need only enforce this requirement to regain security. However, PIN veri(cid:12)cation methods
that support proprietary decimalisation tables are harder to (cid:12)x. A checking procedure
that ensures a mapping of the input combinations to the maximum number of possible
output combinations will protect against the (cid:12)rst two decimalisation table attacks, but
not against the attack which exploits the PIN o(cid:11)set and uses only minor modi(cid:12)cations to
the genuine decimalisation table. To regain full security, the decimalisation table input
must be cryptographically protected so that only authorised tables can be used.
The only short-term alternative to the measures above is to use more advanced in-
trusion detection measures, and it seems that the long term message is clear: continuing
to support decimalisation tables is not a robust approach to PIN veri(cid:12)cation. Unskewed
randomly generated PINs stored encrypted in an online database such as are already used
in some banks are signi(cid:12)cantly more secure.
7 Conclusions
WearecurrentlystartingdiscussionswithHSMmanufacturerswithregardtothepractical
implications of the attacks. It is very costly to modify the software which interacts with
HSMs, andwhile updateof theHSMsoftware is cheaper, thesystemwill still needtesting,
and the update may involve a costly re-initialisation phase. Straightforward validity
checking for decimalisation tables should be easy to implement, but full protection that
retains compatibility with existing mainframe software will be hard to achieve. It will
depend upon the intrusion detection capabilities o(cid:11)ered by each particular manufacturer.
We hope to have a full understanding of the impact of these attacks and of the optimal
preventative measures in the near future.
Although HSMs have existed for two decades, formal study of their security APIs is
still in its infancy. Previous work by one of the authors [5, 4] has uncovered a whole
host of diverse (cid:13)aws in APIs, some at the protocol level, some exploiting properties of
the underlying crypto algorithms, and some exploiting poor design of procedural controls.
The techniques behind the decimalisation table attacks do not just add another string
to the bow of the attacker { they further con(cid:12)rm that designing security APIs is one
of the toughest challenges facing the security community. It is hard to see how any
one methodology for gaining assurance of correctness can provide worthwhile guarantees,
given the diversity of attacks at the API level. More research is needed into methods
for API analysis, but for the time being we may have to concede that writing correct
API speci(cid:12)cations is as hard as writing correct code, and enter the traditional arms race
between attack and defence that so many software products have to (cid:12)ght.
13

Acknowledgements
WewouldliketothankRichardClaytonandRossAndersonfortheirhelpfulcontributions
and advice. Mike Bond was able to conduct the research thanks to the funding received
fromtheUKEngineeringandPhysicalResearchCouncil(EPSRC)andMarconiplc. Piotr
Zielin(cid:19)ski was supported by a Cambridge Overseas Trust Scholarship combined with an
ORS Award, as well as by a Thaddeus Mann Studentship from Trinity Hall College.
References
[1] R. Anderson: Why Cryptosystems Fail Communications of the ACM, 37(11), pp32{
40 (Nov 1994)
[2] R. Anderson: The Correctness of Crypto Transaction Sets Proc. Cambridge Security
Protocols Workshop 2000 LNCS 2133, Springer-Verlag, pp 125{127 (2000)
[3] A. Biryukov, A. Shamir, D. Wagner Real Time Cryptanalysis of A5/1 on a PC
Proceedings of Fast Software Encryption 2000
[4] M. Bond, R. Anderson API-Level Attacks on Embedded Systems IEEE Computer
Magazine, October 2001, pp 67{75
[5] M. Bond: Attacks on Cryptoprocessor Transaction Sets Proc. Workshop Crypto-
graphic Hardware and Embedded Systems (CHES 2001),LNCS2162,Springer-Verlag,
pp 220{234 (2001)
[6] IBM Inc.: IBM 4758 PCI Cryptographic Coprocessor CCA Basic Services Reference
and Guide for the IBM 4758-001, Release 1.31. IBM, Armonk, N.Y. (1999)
http://www.ibm.com/security/cryptocards/bscsvc02.pdf
14