participants do not trust each other, yet are able to somewhat function
through pure magic of modern cryptography [3, 9] rather than through
trusting some financial institutions. Cryptography moves from the status
of disabling some attacks, fixing the security shortcomings of current IT,
to the status of enabler. Making possible new forms of distributed coop-
erative business and services. Cryptography is a disruptive technology.
1.1
Bitcoin and Key Management
In this paper we look at the question of key management in bitcoin and
other crypto currencies. This question actually becomes important for
interesting technical reasons.
1. First, bitcoin offers pseudonymity but poor anonymity, cf. [18, 13].
Using multiple keys helps to achieve better anonymity without really
being a panacea.
2. In addition bitcoin uses a non-orthodox elliptic curve, while in most
cases public keys are only revealed at the first spending transaction.
Therefore it is recommended never to use the same key twice, in
this way the attacker cannot know the public key in advance which
makes key recovery attacks truly improbable; the attacker would
need to compute an Elliptic Curve Discrete Log in matter of a few
seconds cf. [10] in order to have a reasonable chance to double spend
an existing quantity of bitcoins which was already spent.
3. The questions of key management are important in order to insure
reliable backup of bitcoin wallets. If a wallet contains randomly gen-
erated keys, sooner or later it runs out of keys, and the backup does
no longer allow to recover all the private keys and some bitcoins
could be lost. The only solution known is to derive all keys in a
deterministic way.
4. There is also a question of cold storage of bitcoins, for example in
a safe. Using random keys leads to large volumes of keys which are
very hard to manage in practice for humans. Companies which do
cold storage at a large scale for numerous customers need again use
deterministic key management solutions. It appears that billions of
dollars are currently stored in such systems.
1.2
Holy Grail Bitcoin of Key Management: Audit
Capability
In the historical process of development of deterministic key manage-
ment solutions in bitcoin, we have seen another “cryptographer’s dream”
in operation. The idea that it could be possible to assure some level of
transparency in bitcoin world without compromising security. The key
question is the audit capability. For example in a financial institution
or broker firm we would like to give some people, auditors, the capacity
to see all the transactions without being able to spend the money. Simi-
larly in a cold storage system we we would like to know how much money
we have at any moment. this capability is also essential in e-commerce
scenarios, where the server connected to the Internet which accepts pay-
ments must be able to generate public keys on the fly, cf. see Section 2.
This audit capacity is actually the source of all the trouble and the source
of most attacks on this paper. The bottom line is that this “audit-ability
dream” is somewhat possible with current public key cryptography, and
yet somewhat deficient, leading to numerous interesting attacks.
This paper is organized as follows. In Section 2 we study the question of
deterministic private key management. In Section 3 we look at generic
solutions which allow to derive private and public keys with the audit
capability. In Sec. 3.1 we look at a specific solutions which are specified in
BIP032 and in Sec. 3.3 we describe the basic privilege escalation attack.
In Section 4 we discuss the traditional bad RNG attacks on ECDSA,
show that such events have occurred in bitcoin, and introduce more so-
phisticated related random events. In Section 5 we describe new more
general related random attacks.
In Section 6 we describe three advanced combination attacks. In Section
7 we describe three further attacks across two distinct domains which
have no shared secrets. In Section 8 we show that these lead also to
some 6 more analogous attacks in which instead of identical randoms,
related randoms are generated by different users, and also describe a
more general attack with an arbitrary number of domains. Finally in
Section 9 we show another practical attack scenario which takes it even
one stop further: that similar attacks also work in a variety of practical
scenarios where less bad or related randoms occur, however some events
additional in which the private keys overlap have also been detected.
2
Deterministic Private Key Management
In symmetric cryptography we use deterministic key derivation functions
for a number of reasons.
1. They limit key exposure and they limit amount of data that the
attacker can dispose of, for both traditional algorithmic and side-
channel attacks.
2. It avoids expensive storage of too many keys, avoids large databases
and the problem of updating them as new devices join the system.
Now for any system with public key cryptography, we can derive private
keys in the same way, and then derive public keys from private keys on
an individual basis, key per key. We call it Type 1 Wallets in bitcoin
jargon.
Fig. 1. The principle of Type 1 wallets.
Many bitcoin applications use Type 1 solutions and their widespread ap-
plication is attributed to Greg Maxwell, Pieter Wuille and other bitcoin
developers [23, 22, 16]. Type 1 solutions typically derive keys by applying
SHA256 to a master seed with an added counter n.
Type 1 Wallets are secure (if SHA256 is secure) but not very practical.
We would like to have audit capability: the capacity for auditors to see
public keys but not private keys without again having all the problems
implied by a large database of keys. This capacity is also needed in an
e-commerce scenarios we would like a web site to be able to generate
unique keys for receiving bitcoins for a large number of different transac-
tions without being able to spend any of the bitcoin funds (web servers
are a big target for hackers). It also increases anonymity of commer-
cial transactions knowing that the bitcoin blockchain is public (web sites
which accept credit card payments do not have this problem).
2.1
Type 2 Key Management Solutions
The crucial property sought for a wallet to be a Type 2 solution is as
follows. We want to have 2 derivation functions such that the following
diagram commutes. For example P K
3
should be the same whether it is
computed by the Public derivation function or if it is computed from
SK
3
. In addition the public seed should NOT reveal the private seed or
any of the private keys. All the 4 derivations on the diagram should be
done with one-way functions.
Fig. 2. The principle of Type 2 wallets: the two key derivation functions must be such
that the diagram commutes and public keys P K
i
obtained either way are the same.
So far it is possible to see that the goal is achievable. We can propose
solutions which satisfy these requirements. The main idea [attributed to
Greg Maxwell and Peter Wuille] is as follows, cf. [23, 22, 16]. Both deriva-
tion functions are essentially the same, however in the ”Private Type 2
Derivation” function the private key is simply used to compute the cor-
responding public key. We will use two distinct algebraic properties of
ECDSA keys in order to be able to obtain two such solutions for ECDSA
keys specifically. In fact more advanced solutions exist which have also
two additional properties:
1. They have additional long secrets beyond public/private keys. These
additional secrets are called ”chain” values, so that knowing all
ECDSA keys still does not allow to derive further keys.
2. They work across several levels of further specializatiin/derivation
(cf. Fig. 3, below), allowing very interesting and elegant key man-
agement solutions with multiple hierarchical security domains and
sub-domains.
3. Similarly they allow to generate individual keys for individual trans-
actions.
Fig. 3. The principle of hierarchical Type 2 wallets across several levels: the diagram
should commute with all possible paths giving the same result.
Again it appears that the goal is achievable and we are going to describe
some generic and one more specific solutions which seem to satisfy these
requirements. We are going to need the following definitions:
Definition 21. We call a security sub-domain a set of all the nodes
in the DIRECTED graph as represented in Fig. 3, with all the nodes
and leaves BELOW one point/edge, in the sense of the arrows, which
corresponds to what can be computed: lower level keys can be computed
from higher level, and public seeds CAN be computed from private seeds.
Remark: The level of access on any user in an HD wallet system can be
seen as a union of several security domains in the sense of Def. 21 above.
However the possession of keys which give access to several domains can
have unintended consequences, as a clever attacker will be sometimes able
to gain access to additional security domains. See our Privilege Escalation
Attack in Prop. 32 on page 11.
3
Generic Solutions
In this section we two generic solutions which works across an arbitrary
number of levels. In this paper we only cover solutions with additional
secrets (chain variables) so that seeds in Fig. 3 will now always be of
specific form asymmetric key (left part) concatenated with a Chain vari-
able C (right part). Both solutions are specific to ECDSA in the sense
that a similar solution probably wouldn’t exist if a different digital sig-
nature scheme with less special algebraic/homorphic properties was used
in bitcoin. It is possible to see that both solutions could also work with
traditional DSA-like signatures. In this paper we will use the additive
notation (as in ECDSA) and ignore their DSA versions. All attacks in
this paper are totally independent of the elliptic curve used.
We are going to need the following definition:
Definition 31. We assume a public key cryptosystem such that the pri-
vate key can be derived in an efficient deterministic way from the public
key (e.g. ECDSA). We call an Extended Private Key a concatenation of
a private key and an additional secret x
C
which will be called Chain
variable.
We call an Extended Public Key a concatenation of a public key and an
additional secret x
C
which under right circumstances it will be the same
value as above.
We call Neutering or N the operation which transforms an Extended
Private Key into an Extended Public Key as follows: we derive the public
key for the left part by EC base point multiplication, and we copy the
Chain variable x
C
, so we have
N (k||x
C
) = (k.G||x
C
).
This operation N of Neutering is what is need to make diagrams such
as in Fig. 3 commute. In all the solutions presented in this paper and
on all pictures in this paper each of the horizontal arrows represents an
application of N on extended keys with chain values.
3.1
Two Solutions
Interestingly, even though the problem is not completely trivial to solve,
there two distinct major generic solutions to this problem known, both
of which allow flexible audit capabilities. By flexible we mean that we
can give auditors extended keys (cf. Def. 31) which allow audit of all
operations in ANY security sub-domain as defined in Def. 21 above, and
which works down across an arbitrary number of levels. We present both
solutions in Fig. 4 and Fig. 5 below.
Fig. 4. A generic Solution1 using the EC scalar multiplication property, HM is an
arbitrary OW function with 2 inputs called M and K. In practice HMAC can be used.
Fig. 5. A generic Solution2 using the EC homomorphic property, HM is an arbitrary
OW function with 2 inputs called M and K. In practice HMAC can be used.
In all cases we denote by q be the order of the EC subgroup used. In
bitcoin q = 2
256
− 2
32
− 977 we have. In both solutions if we assume
that we apply N to the inputs on the left side and we will get the inputs
on the right side, the same relationship also holds for the two outputs,
i.e. we can apply N to the outputs on the left side and we will get the
same correct outputs on the right side. We refer to Proposition 31 and
Corollary 31 below for a proof of correctness and applicability of these
two solutions across an arbitrary number of levels as in Fig. 3. We have
the following results:
Proposition 31. For both Solution1 and Solution2 with notations as
in Fig. 4 and Fig. 5 if we have that
N (k||x
C
) = (K||x
C
).
which means that K = k.G and x
chain
is the same on both sides, THEN
then we obtain the right key with
N (k
i
, c
i
) = (K
i
, c
i
).
or in other words we have K
i
= k
i
.G and the same c
i
is obtained on
both sides.
Proof: The proof is obvious by following the identical values on the di-
agram in Fig. 4 and Fig. 5. In Solution1 in Fig. 4 the public key K
i
on
the right hand side is the correct key for k
i
on the left because we have
k
i
.G = (y
Li
∗ k
L
).G = y
Li
.(k
L
.G) = y
Li
.K
L
= K
i
.
In Solution2 in Fig. 5 and in BIP032 in Fig. 6 the public key K
i
on the
right hand side is the correct key for k
i
on the left because we have
k
i
.G = (y
Li
+ k
L
).G = y
Li
.G + k
L
.G = y
Li
.G + K
L
= K
i
.
More generally for both solutions and by induction it is easy to see that:
Corollary 31. Both the solution of Fig. 4 and Fig. 5 can produce com-
muting diagrams across an arbitrary number of levels such as in Fig. 3
in which all directed paths produce exactly the same public keys.
Remark 1. Both solutions exploit two distinct properties on ECDSA
keys. We have only two major solutions in Fig. 4/5 and it is NOT easy
to produce similar solutions. It is not true that one could use an arbitrary
group operation to combine y
L
with k and produce other solutions
4
.
Remark 2. Both solutions have been implemented and are very widely
used in bitcoin community; in bitcoin exchanges, cold storage systems,
wallet systems, electronic commerce solutions, trading solutions etc. The
Solution1 is based on the initial proposal which has been used in vari-
ous systems such as Electrum before it was changed on 30 April 2013 to
implement Solution2 which is now standardized as part of BIP032 spec-
ification and more widely used. It is not clear that Solution2 was ever
meant to be more secure than Solution1, however it was claimed to be
faster and easier to implement [22]. In this paper was show that Solu-
tion2/BIP032 is NOT more secure than Solution1, and we show similar
sort of attacks on both solutions.
4
Quite possibly there would be NO solution at all if bitcoin has used a different sort
of digital signature scheme (e.g. Lamport signatures [6]) which does not have special
algebraic properties.
3.2
A Specific Solution: Details About BIP032
BIP032 is a specific application of Solution2 based on HMAC-SHA512.
Fig. 6. Details of BIP032 public and private key derivation mechanisms known as CDK
or Child Key Derivation functions, both work together.
Fig. 7. The ”Hardened” version of BIP032 private key derivation function when i ≥
2
31
. This is version is back again a Type1 solution like on Fig. 1. Here the ”compatible”
public version does not exist and there is no way to derive public keys in bulk.
3.3
Basic Attacks Against The Two Solutions and
BIP032
Both solutions admit a number of relatively simple attacks. These attacks
can be considered as known and most have already been discussed in
various bitcoin forums for at least one of the two versions, and are also
mentioned by the author of BIP032 Pieter Wuille in both [22] and [23].
However there is no systematic study of these attacks and it was not
clear if similar attacks always exist for both solutions and if they work
across an arbitrary number of levels. All these attacks can be summarized
in one single general privilege escalation attack. This is what we are to
show now.
Fig. 8. One high-level extended public key and ANY private key at any level below
combined allow to compromise the master private secret k and derive any other key.
Proposition 32 (Privilege Escalation Attack). For both solutions
in Fig. 4/5 applied across an arbitrary number of levels we have: if the
attacker disposes of the audit key K, x
c
for one Security Domain as
defined in Def. 21 AND if the attacker knows just one private key at any
level one or several degrees below k, say he knows a private key k
I
where
I is a a set of 1 or more indices, THEN the attacker can compute the
high level private key k at the level of the audit key known.
Therefore he can also compute any other private key and corresponding
secret chain value k
J
, c
J
for any other arbitrary path J of length ≥ 1
below k and similarly also arbitrary public keys below K.
Proof: For simplicity we first assume there is only one layer of key deriva-
tion functions. We recall that q is the order of the EC subgroup used.
With Solution1 one private key is of the form y
1
∗ k mod q and the other
is y
2
∗ k mod q. The auditor key K, x
c
allows to compute both y
1
, y
2
so
he can compute k and y
2
∗ k mod q.
With Solution2 or/and BIP032 one private key is of the form y
1
+ k
mod q and the other is y
2
+ k mod q. The auditor key K, x
c
allows to
compute both y
1
, y
2
so he can compute k and y
2
+ k mod q. By induc-
tion we can go up an arbitrary number of levels as long as we have the
audit extended key on these levels which allow us to compute all the y
i
values needed.
Discussion. The privilege escalation attack of Prop. 32 above can be
considered as a very serious attack, because a coalition of two people can
compromise the whole system. However if all the digital signatures are
done on the cold machines and in absence of an exploit or malware on
the cold machine, the system remains secure. In this paper we describe
many new attacks all of which operate under a weaker assumption and
are such that the attacker does not need to run any code on the cold
machine, he just needs to identify some bad randomness events.
4
Bad Randomness Attacks On Bitcoin
In this Section we summarize what is known about bad randomness
attacks in bitcoin. In fact the full picture about these vulnerabilities
have never been made public and it this section we disclose a number
of facts and the relevant security and impact analysis which has not yet
been published.
4.1
History Of Bad Randomness Attacks on Bitcoin
On 28 January 2013 Nils Schneider have reported that repeated randoms
have occurred in the bitcoin data. He has found two bitcoin transactions
which use the same random and published this fact on his blog [19]. The
problematic transaction in question was from March 2012 and there-
fore the problem remained undetected for quite some time. Few days
later Moore and Wustrow publish more detailed results [14]. They have
examined some 11 million bitcoin transactions to date and found 126
repeated random values on 256 bits which appear in 316 different trans-
actions. They found that most values on 256 bits were repeated only
once. However some values were repeated a few times, and the value
D47CE4C025C35EC440BC81D99834A624 875161A26BF56EF7 FDC0
F5D52F843AD1 which was already found by Schneider [19] was repeated
more than 50 times In August 2013 we found on the Internet another file
posted anonymously by a certain Greg, which contained 131 bad ran-
doms. However already at this moment this file already was incomplete,
we have found a few more such events at the time and we have reported
some additional events inside the Code Breakers group at LinkedIn. At
30C3 conference in Germany on 28 Dec 2013 [5] Nadia Heninger have
reported that they have identified a bitcoin user on the blockchain which
has stolen some 59 BTC due to these bad randomness events which
address is 1HKywxiL4JziqXrzLKhmB6a74ma6kxbSDj, see also [21]. To
this day this money remains on the blockchain and has not been spent.
4.2
A Recent Outbreak of Bad Random Events
Most of the events form 2013 have happened in May 2013, cf. Fig. 9,
and then have almost entirely disappeared, presumably because the An-
roid applications affected by the problem have been reportedly updated
(including our own) and applications have transferred bitcoins automat-
ically to new accounts with fresh bitcoin addresses.
A recent massive outbreak of bad random events have occurred in May
2014, cf. [21] and Fig. 9 below.
Here is an example of a very recent bad random from 28 September 2014:
0f25a7cc9e76ef38c0feadcfa5550c173d845ce36e16bde09829a3af57097240.
This random r appears 8 times in block 322925 and has been used with
different private keys so that it is not obvious if bitcoins can be stolen,
see Section 4.3 below. However if any of the HD Wallet solutions have
been used to generate at least two of these keys, for sure bitcoins can be
stolen, and not only from these 8 accounts, but from potentially a larger
domain with a lot more accounts, cf. our later Proposition 63.
Fig. 9. Bad randomness events in bitcoin blockchain, the sending address is represented
on the vertical axis. Short horizontal lines suggest that some vulnerable applications
have used the same address many times in a row, vertical lines suggest that other
applications have used many randomly generated addresses.
4.3
Basic Attacks with Bad Randomness and ECDSA
In order to explain better the events, we need to explain what are the
exact implications of using the same random twice in a DSA/ECDSA-
like signature. Bad randomness in various industrial systems is a widely
known problem, see [7, 12, 11, 19, 21]. In this paper we need first to recall
basic attacks more or less in the line of specific incidents already stud-
ied in previous works [1, 11, 19, 21] and later we will propose new more
advanced attacks. We recall that:
Definition 41 (ECDSA Signature).
If the integer 0 < k < q − 1 is the ECDSA private key, and q is the
order of the base point G, an ECDSA signature is computed as follows.
We pick a random non-zero number a mod q and the signature of m is
the pair r, s with
r = (a.G)
x
; s = (H(m) + kr)/a
mod q
It is easy to see that we have the following two results:
Proposition 41 (Random Recovery in ECDSA). If the attacker
knows the original random a, one message and one valid signature m, r, s,
he can compute the private key k with the following formula:
k = (sa − H(m))/r
mod q.
Proposition 42 (Key recovery From Random). From one valid
signature m, r, s and the private key k one can later recover the original
random a as follows:
a = (H(m) + kr)/s
mod q.
Therefore:
Proposition 43 (Attack on ECDSA with Bad RNG). If the same
random a is used in two different ECDSA signatures m
1
, r, s
1
, and m
2
, r, s
2
,
each of the two signers can compute the private key of the other signer.
Proof: We recover a due to Prop. 41 and then we apply Prop. 41.
Remark 1. It is possible to see that for the public (people other than
the 2 signers) and from the public data m
1
, r, s
1
, and m
2
, r, s
2
alone, it
is NOT possible to compromise the private keys of the two users
Remark 2. This changes if we know the source code of the RNG and
have a chance to reproduce a. Then we can recover both private keys (cf.
Prop 42).
Remark 3. In bitcoin network the compromise of one private key allows
to steal all the money in one specific account. They can also be stolen
at any moment in the future, i.e. moneys which were not there earlier.
Money can be stolen even the RNG has been fixed and produces good
random numbers later on, however some senders still use the old public
bitcoin address. This is very hard to avoid: potentially anyone can send
money to an older address.
4.4
Special Cases
There is an interesting special case.
Proposition 44 (Bad RNG ECDSA Same User Attack). If two
digital signatures are produced with the same private key AND the same
random, then given m
1
, r, s
1
, and m
2
, r, s
2
, anyone can compute the
private key of that person at any moment.
Proof: We recall how it is done (cf. also [19]). We have
(s
1
a − H(m
1
)) = ak = (s
2
a − H(m
2
))
mod q
This gives a = (H(m
1
) − H(m
2
))/(s
1
− s
2
) mod q and we can recover
the private key k using Prop. 42.
There are also cases in which the keys acan be recovered by anyone,
basically if such events happen many times. For example we have:
Proposition 45 (Attack with Two Bad Random Events). If two
identical randoms a and b are used just once by two different users (each
user needs to use both a and b, then anyone can recover the private keys
for both users.
Proof: This is a special case of our later Prop. 71.
4.5
Our Bitcoin Network Scan - Classification of Events
In our scan there are three basic sorts of events:
1. The same random was used by the same person with the same private
key in the same transaction. In this case following Prop. 44 above
anyone can compute the private key of that person at any moment
and steal his coins (also in the future). This seems unlikely but has
already happened in the original case from March 2012 cf. [19] and
is also the case for the majority of 148 events.
2. The same random is used by two different people. In this case fol-
lowing Prop. 43 above each signer can compute the private key of
the other signer and steal all his coins, but there is no immediate
5
threat of other people stealing bitcoins.
3. There is a mix of 1. and 2. The same random has been used at least
3 times once it was used twice by the same person, and sometimes
it was also used by other people. Then Prop. 44 above the random
a can be recovered and then following Prop. 42 the private keys for
the other users can also be compromised.
4. The same random was used by the same person with different private
keys, which is the same as case 1. above however that person already
knows all the keys and there is no harm.
The crucial remark is that in case 2. attackers who are not any of the two
people can do nothing. In this paper we present a number of combination
attacks where this will nevertheless work and lead to a compromise of a
whole large security domain (all private keys will be recovered).
4.6
Which Applications Are Vulnerable
Some analysis of which applications are vulnerable can be found at the
end of [21]. The analysis is very disturbing: it appears that the bitcoin
core client has NOT yet finalized the process of several security upgrades
which propose a solution to this problem (deterministic random genera-
tion, cf. [17, 21]) event though a first solution was already submitted in
January 2013. It appears also that Blockchain.info wallet is also unsafe
as of today.
5
This will change if this sort of thing happens more than once, see for example Prop.
45 above and also point 3. and also cf. advanced combined attacks in Section 6 and
7 below.
5
Advanced Attacks With Related Randoms
In the well known research on bad RNG in public key cryptography and
TLS connections [11] there are two sorts of events:
1. Events when two primes generated by two different users are iden-
tical. Thousands of such cases have been identified in [11] and we
have identified a few hundreds more such events in our own scan.
2. Events where two primes are NOT identical but there are shifted by
a number of bytes.
To the best of our knowledge such events 2. were NOT yet identified in
the public literature in TLS connections. We expect that a very large
number of such events have occurred in the real life. This is based on our
own experimentations done at UCL: recovering some 8000 primes which
have been obtained by our students in the same way as in [11]. We have
then discovered that 10 of these primes had at least 200 bits in common
with another prime in the set with various shifts, cf. [1].
This is of course just a tip of the iceberg: there are many other primes in
this dataset which share long strings of bits with other primes, and we
have NOT yet been able to discover them because they belong to primes
which have not yet been factored with the method of [11]. Some related
research can be found in [4].
5.1
Related Randoms In Bitcoin
What happens if two related randoms are generated by two different
bitcoin clients? Again depending on how big is the overlap this is going to
allow each of these people to steal other people’s bitcoins. Let us assume
that two random a and b are such that b starts at a position 8 bits later
with the same RNG setup. again this sort of event has happened in the
real life. Then both randoms a and b can be recovered as follows:
1. We have:
b + c · 2
256−8
= 256 · a + d
Where c and d are very small 8-bit integers with can be guessed
2. Therefore if r
1
, s
1
is a digital signature obtained with a, and r
2
, s
2
is the one obtained by another user with b, we have on the elliptic
curve:
r
2
+ c · 2
256−8
.G = 256 · r
1
+ c.G
3. This equation allows to find the right pair c, d in time of 2
16
EC
operations.
Again as with Prop. 43, in principle this does not YET allow anyone
to steal bitcoins of these two users. For now only these two users can
steal some bitcoins. This will however change if HD Wallets with any
Solution1 or 2 are also used at the same time, see Section 8 and 9 below.
Remark. In a future release of this paper we are going to reveal a much
more efficient and more practical method for finding such relations.
6
Advanced Combination Attacks
This is the main part of this paper. It contains several new attacks never
published before and never claimed before. The new attacks are com-
bination attacks which combine distinct vulnerabilities we has studied
so far (cf. Section 3.3 and Section 4). They allow to recover private
keys which none of the two vulnerabilities alone would allow to hope to
recover. In fact typically it can lead to a total compromise of a given
system (all private key revealed), or to a partial compromise (a whole
sub-domain). We describe several such attacks which work across an ar-
bitrary number of levels of key derivation and in a variety of practical
scenarios. The first one is going to be pretty trivial.
Proposition 61 (Combined Attack Bad RNG+Solution1/2). If
the attacker knows the Extended Public Key K||x
C
for a whole system or
for any security sub-domain as defined in Def. 21 AND if two identical
randoms were used by a single user anywhere in the system, then he
can recover ALL private keys in this sub-domain as well as the master
private key for the same whole sub-domain.
Proof: We combine Prop. 44 page 15 and Prop. 32 page 11.
This was easy, however we cannot hope that such events will happen.
What if the same random is used by two different users? This is more
probable. Here we will have a more serious attack which will really show
how fragile HD Wallet solutions are.
Proposition 62 (Combined Attack On Solution1). If the attacker
knows the Extended Public Key K||x
C
for a whole system or for any
security sub-domain as defined in Def. 21 AND if two identical randoms
are used just once by any pair of users in this domain, and if the relative
paths I
1
and I
2
of these users in the domain are known
6
, then he can
recover ALL private keys in the domain as well as the master private key
for the same whole sub-domain (or for the whole system).
Proof: We proceed as follows. From K||x
C
and I
1
the attacker can com-
pute all the multipliers y
L
which are applied to the private key along
the path in one or more steps such as in Fig. 4. Let y
1
be the product
of these multipliers. Similarly let y
2
be the same quantity for the second
path I
2
. Then the two private keys are ky
1
and ky
2
with unknown k. We
have:
as
1
= (H(m
1
) + rky
1
)
mod q
as
2
= (H(m
2
) + rky
2
)
mod q
(H(m
1
) + rky
1
)s
2
= (H(m
2
) + rky
2
)s
1
mod q
k(ry
1
s
2
− ry
2
s
1
) = H(m
2
)s
1
− H(m
1
)s
2
mod q
k =
H(m
2
)s
1
− H(m
1
)s
2
ry
1
s
2
− ry
2
s
1
mod q
6
As in Section 3.3 I can be any sequence of 1 or more indices, for example 3/5, from
the unknown Extended Private Key k||x
C
The last equation allows to recover the top-level private key k for this
domain. This completes the proof.
We have a similar attack for Solution2 of Fig. 5 BIP032 in particular.
Proposition 63 (Combined Attack On Solution2/BIP032). If the
attacker knows the Extended Public Key K||x
C
for a whole system or for
any security sub-domain as defined in Def. 21 AND if two identical ran-
doms are used just once by any pair of users in this domain, and if the
relative paths I
1
and I
2
of these users in the domain are known then he
can recover ALL private keys in the domain as well as the master private
key for the same whole sub-domain (or for the whole system).
Proof: Again From K||x
C
and I
1
the attacker can compute all the dif-
ferences y
L
mod q which are applied to the private keys along the path
as in Fig. 5. Let y
1
be the sum of these differences
mod q and let y
2
be the same sum quantity for the second path I
2
. Then the two private
keys are k + y
1
and k + y
2
with unknown k. We have:
as
1
= (H(m
1
) + rk + ry
1
)
mod q
as
2
= (H(m
2
) + rk + ry
2
)
mod q
(H(m
1
) + rk + ry
1
)s
2
= (H(m
2
) + rk + ry
2
)s
1
mod q
kr(s
2
− s
1
) = H(m
2
)s
1
− H(m
1
)s
2
− r(y
1
s
2
− y
2
s
1
)
mod q
k =
H(m
2
)s
1
− H(m
1
)s
2
− r(y
1
s
2
− y
2
s
1
)
r(s
2
− s
1
)
mod q
Again the last equation allows to recover the top-level private key k.
7
Advanced Combination Attacks Across
Domains
In this section we study even more advanced attacks. In these attacks we
will attack two totally independent HD wallet key management solutions.
For example that there are two bitcoin exchanges with cold storage. If by
unhappy accident a certain random is used twice in one single domain,
we already know that this domain is compromised. Now what happens if
the same random is used twice but in two different domains. In principle
there is no hope for any attack. Now let us assume that it happens twice
with two different unknown randoms a and b. Then there is an attack!
First we look at Solution1. We have
Proposition 71 (Combined Attack Solution1 Two Domains). If
the attacker knows the audit keys from two different domains with So-
lution1 AND if two identical randoms a and b are used just once by one
user in each domain (four users total) and if we know relative paths of all
users, then the attacker can recover ALL private keys and both master
private keys in both domains.
Proof: Here the four private keys are ky
1
, k
0
y
2
, ky
3
, k
0
y
4
, with unknown
k and unknown k
0
but the four values y
i
can be computed from the audit
keys for known paths. We have:
as
1
= (H(m
1
) + r
a
ky
1
)
mod q
as
2
= (H(m
2
) + r
a
k
0
y
2
)
mod q
bs
3
= (H(m
3
) + r
b
ky
3
)
mod q
bs
4
= (H(m
4
) + r
b
k
0
y
4
)
mod q
ky
1
s
2
− k
0
y
2
s
1
=
H(m
2
)s
1
− H(m
1
)s
2
r
a
mod q
ky
3
s
4
− k
0
y
4
s
3
=
H(m
4
)s
3
− H(m
3
)s
4
r
b
mod q
This gives us 2 linear equations with 2 unknowns k and k
0
which allows to
compute BOTH master private keys(!). All the other values are known.
A similar attack holds for Solution2/BIP032 but the computations are
different.
Proposition 72 (Combined Attack Solution2/BIP032 Two Do-
mains). If the attacker knows the audit keys from two different domains
with Solution2 AND if two identical randoms a and b are used just once
by one user in each domain (four users total) and if we know relative
paths of all users, then the attacker can recover ALL private keys and
both master private keys in both domains.
Proof: Here the four private keys are k + y
1
, k
0
+ y
2
, k + y
3
, k
0
+ y
4
, with
unknown k and unknown k
0
but the four values y
i
can be computed from
the audit keys for known paths. We have:
as
1
= (H(m
1
) + r
a
k + r
a
y
1
)
mod q
as
2
= (H(m
2
) + r
a
k
0
+ r
a
y
2
)
mod q
bs
3
= (H(m
3
) + r
b
k + r
b
y
3
)
mod q
bs
4
= (H(m
4
) + r
b
k
0
+ r
b
y
4
)
mod q
ks
2
− k
0
s
1
=
H(m
2
)s
1
− H(m
1
)s
2
r
a
− y
1
s
2
+ y
2
s
1
mod q
ks
4
− k
0
s
3
=
H(m
4
)s
3
− H(m
3
)s
4
r
b
− y
3
s
4
+ y
4
s
3
mod q
This also gives us 2 linear equations with 2 unknowns k and k
0
which
allows to compute BOTH master private keys(!).
7.1
Heterogenous Domains
Furthermore:
Proposition 73 (Attack Across Domains Using Different Key
Management Solutions). A similar attack exists if one bitcoin key
management system uses Solution1 and another uses Solution2 or BIP032.
Proof: Here is how all the private keys can be recovered in this case:
as
1
= (H(m
1
) + r
a
ky
1
)
mod q
as
2
= (H(m
2
) + r
a
k
0
+ r
a
y
2
)
mod q
bs
3
= (H(m
3
) + r
b
ky
3
)
mod q
bs
4
= (H(m
4
) + r
b
k
0
+ r
b
y
4
)
mod q
ky
1
s
2
− k
0
s
1
=
H(m
2
)s
1
− H(m
1
)s
2
r
a
+ y
2
s
1
mod q
ky
3
s
4
− k
0
s
3
=
H(m
4
)s
3
− H(m
3
)s
4
r
b
+ y
4
s
3
mod q
Again both master private keys can be determined from the last 2 equa-
tions which will compromise all the keys in both domains.
7.2
Discussion, Hardened Option In BIP032
It is important to recall that in all the cross-domain attacks in this
Section 6 the different domains do NOT initially share any keys, yet at
the end ALL keys in ALL the domains are recovered.
Interestingly the BIP032 key management standard has an interesting
option of ”hardened” key derivation function which is applied if any index
i > 2
31
. If we apply these key derivation functions we obtain distinct
domains with independent extended public keys, a near-equivalent of
truly independent
7
domains. then we just treat these as distinct domains
and all the three attacks in this Section 6 still do apply. Therefore it is
possible to claim that the hardened option in BIP032 is NOT yet an
ideal solution against the attacks described in this paper. Moreover the
hardened option options makes key management against more difficult
and makes it necessary for auditors to hold many more keys.
7.3
Future Research
Many more attacks such as described in this paper exist. It is possible
to see that if we have an arbitrary number of bitcoin exchanges or cold
storage solutions each of which uses an independent master seed, if we
dispose of all the audit keys (i.e. Extended Public Keys for all the ex-
changes, and if repeated randoms can be obtained in operation for some
pairs of exchanges, and if some sub-branches of the BIP032 system are
hardened, after a small number of events of this type the master keys of
ALL domains can be computed and all bitcoins in ALL the exchanges
can be stolen.
7
Their private keys are still related BUT attacks such as Privilege Escalation Attack
of Prop. 32 are no longer possible.
8
Generalized Combined Attacks with Related
Randoms
There are many further improvement attacks which will NOT by imme-
diately visible on the blockchain. Basically it is possible to see that all
the 3 previous attacks in Section 6 and the 3 further attacks in Section
7 can also be generalized to work without a single repeated random.
Instead bad randoms need to be shifted by a small shift, and such pairs
of related bad randoms can be recovered in an offline stage as described
in Section 5.1 from the public keys. Then after the offline stage, we
obtain explicit linear equations with known coefficients c and d such as
in 5.1 on the private keys. With these linear relations we can apply all
the 5 attacks form Section 6 and their generalizations to more distinct
security domains, and each time we will need less linear equations such
as produced in the last two formulas in Section 7.1, because we already
have additional linear equations. As soon as the rank of our linear system
become full, all the private keys for ALL security domains involved can
be recovered.
Attacks with related random events are particularly scary. With a larger
shift between two randoms a and b which are not visible to the attack, it
becomes more probable that they actually occur in the real life and yet
increasingly harder to detect. We cannot claim that we have found all
such events. It will take month, maybe years an potentially much longer
to detect such security incidents in the public blockchain. Therefor these
attacks are particularly scary. It is possible to see that no cold storage
system using BIP032 and a randomly generator can claim that they are
secure. There is no guarantee that all their bitcoin are NOT stolen today
or tomorrow using one of our attacks. However it is possible to see that
if they apply a deterministic method such as RFC6979, cf. [17, 21], none
of the advanced combination attacks in this paper will work (in contrast
the simpler coalition attack of Prop. 32 will still work).
9
Additional Attacks: Mix of Related Randoms
and Related Keys
The attacks of Section 8 can be taken one stop further under different
assumptions. Similar attacks also work in a variety of practical scenarios
where less bad randoms which relate randoms occur. Then we can also
generate additional linear equations with coefficients which may have
very small entropy on the private keys, and such relations can be effi-
ciently discovered in an offline stage in the same way as in Section 5.1.
Again as above, this leads to even more linear relations on the private
keys which can be recovered in an earlier stage of the attack.
We can mix the equations obtained from identical randoms, those which
come from related randoms, those from related keys, and also simply
those which come from identical keys. All these sorts of events are plau-
sible and have happened hundreds of times in real life situations, some
as recently as September 2014, see Section 4. They have a cumulative
effect. Again we can apply all the 5 attacks form Section 6 and their
generalizations to more distinct security domains and steal all bitcoin
for ALL security domains involved as soon there was a sufficient number
of bad related random events of these 4 distinct sorts.
10
Additional Scenarios with an Insufficient
Number of Problematic Events
Another practical scenario is that both randoms and keys have been gen-
erated using the same RNG with a shift. Then again it is possible to see
that all the same attacks in Section 6, Section 7 and their generalizations
to more distinct domains can be executed.
Finally, it is interesting to understand what happens WHEN in a general
scenario with say 5 security domains and 4 bad random events identified
by the attacker in the blockchain, after long and painful brute force
exploration of plausible relations based on some heuristics, for example
which users are more likely to use the same software setup and use the
same randomness, the attack does NOT yet have enough linear equations
to recover all the private keys. Then our attack still allows at least some
people in the system, who do have additional keys, like ordinary users
8
who will be able to steal other people’s bitcoins not only in the system
in which they belong, but also from other distinct security domains.
8
The may have just one private key for one transaction, or may have recorded the seed
used in their RNG in just one spending transaction, which is sufficient to recover a
private key, cf. Prop. 42.
11
Conclusion
In this paper we have looked at the question of security of current key
management solutions standardized in bitcoin BIP032 and their earlier
variants. It appears that billions of dollars are stored in such systems
worldwide. In such systems private keys are stored on ”cold” machines
which are never connected to the Internet. In practice Type2 HD Wallet
solutions (both current Solution2/BIP032 and another earlier Solution1)
are maybe the only way known to build cold storage, exchange/trading,
electronic commerce and other business systems which are ”manage-
able” in practice. They avoid the explosion in the number of keys and
managing updates if new users/keys are added to the system.
Unhappily such as systems are quite centralized and super fragile. They
have high-level audit keys which are typically stored and highly avail-
able on hot machines or are available for certain people: auditors, servers
which deal with accepting payments e.g. in electronic commerce, etc.
These audit capabilities create a lot mathematical dependencies between
various keys. Now if only there are some more such dependencies due to
other reasons such as bad randomness or other incidents at operation in
some remote corner of the system, everything collapses very badly. Our
attacks allow to recover private keys which none of the earlier attacks in
isolation could hope to recover. In this paper we show that both older
Solution1 and newer Solution2/BIP032 are vulnerable to advanced com-
bination attacks. Classical bad random attacks typically concern only
few users which are very lucky and can steal other people’s bitcoins. In
our new attacks all bitcoins in cold storage over whole large security
domains can be stolen, without executing any exploit or any malicious
code on the cold machine, Our results generalize to an arbitrary number
of distinct security domains. In certain circumstances all bitcoins in cold
storage and at each domain can be compromised. Or certain users in
one domain can help to steal keys in another domain, cf. Section 10.
Attacks with related randoms are particularly scary, cf. Section 8. Such
attacks have not yet been studied in bitcoin literature. They are not
discovered easily, it is more probable that such events occur in the real
life than simpler events we already discovered, and yet they are much
LESS likely to be detected in time, before it is too late.
The impact of this paper on current bitcoin storage systems is hard to
evaluate. On the one hand side we can always hope that there wasn’t
enough bad random events inside the domains using BIP032 to be com-
promised. On the other hand unless they apply a deterministic method
such as RFC6979 from [17, 21], which is a definite solution for most at-
tacks, no existing bitcoin cold storage system using BIP032 should feel
too confident to be secure against the attacks described in this paper.
Open Problems. An interesting question is whether there exist a key
management scheme for ECDSA with flexible audit capabilities such that
the attacks such as described in this paper would not work. If it exists,
the current bitcoin key management standard BIP032 should be replaced
by a new standard.
References
1. E. Alashwali, Cryptographic Vulnerabilities in Real-Life Web Servers,
Master thesis, M.Sc. in Information Security, University College Lon-
don, 31 August 2012.
2. Adam Back: Hashcash - A Denial of Service Counter-Measure, http:
//www.hashcash.org/papers/hashcash.pdf, August 2002.
3. Simon Barber, Xavier Boyen, Elaine Shi, and Ersin Uzun: Bitter to
Better : How to Make Bitcoin a Better Currency, In Financial Cryp-
tography and Data Security, FC’12, Springer, 2012.
4. Daniel
J.
Bernstein,
Yun-An
Chang,
Chen-Mou
Cheng,
Li-
Ping Chou, Nadia Heninger, Tanja Lange, Nicko van Someren:
Factoring
RSA
Keys
from
Certified
Smart
Cards:
Copper-
smith in the Wild, In Asiacrypt 2013, pp. 341-360. Slides:
http://www.hyperelliptic.org/tanja/vortraege/20131030.pdf
5. Daniel Bernstein, Nadia Heninger, Tanja Lange: 2013 - The
Year In Crypto, Presentation at 30th Chaos Communication
Congress [30c3], Congress Centrum Hamburg, Germany, 28 De-
cember 2013 http://www.internet-history.info/media-library/
mediaitem/1829-the-year-in-crypto-30c3.html
6. Vitalik Buterin: Bitcoin Is Not Quantum-Safe, And How We Can Fix
It When Needed, 30 July 2013, http://bitcoinmagazine.com/6021/
bitcoin-is-not-quantum-safe-and-how-we-can-fix/.
7. Nicolas Courtois, Daniel Hulme, Kumail Hussain, Jerzy Gawi-
necki, Marek Grajek: On Bad Randomness and Cloning of Con-
tactless Payment and Building Smart Cards,
In IWCC 2013,
International Workshop on Cyber Crime, San Francisco, USA,
24 May 2013. Slides: http://www.nicolascourtois.com/papers/
buildtransc_rng_sf_22052013.pdf
8. Nicolas T. Courtois, Pinar Emirdag and Daniel A. Nagy: Could Bit-
coin Transactions Be 100x Faster? in proceedings of SECRYPT 2014,
28-30 August 2014, Vienna, Austria.
May 2014,
9. Wei Dai: B-Money Proposal, 1998, http://www.weidai.com/bmoney.
txt
10. Christian Decker, Roger Wattenhofer: Information propagation in
the bitcoin network,
13-th IEEE Conf. on Peer-to-Peer Computing,
2013.
11. N. Heninger, Z. Durumeric, E. Wustrow, and J. A. Halderman, Mind-
ing Your Ps and Q’s: Detection of Widespread Weak Keys in Network
Devices,
2012, https://factorable.net. Accessed: 23-Jul-2012.
12. A. Lenstra, J. Hughes, M. Augier, J. Bos, T. Kleinjung, and
C. Wachter: Ron was wrong, Whit is right, 2012. At http://eprint.
iacr.org/2012/064.pdf.
13. Sarah
Meiklejohn,
Marjori
Pomarole,
Grant
Jordan,
Kirill
Levchenko, Damon McCoy, Geoffrey M. Voelker, and Stefan Savage:
A Fistful of Bitcoins: Characterizing Payments Among Men with No
Names, In IMC 2013.
14. Jonathan
Moore,
Eric
Wustrow:
Bitcoins
and
entropy,
https://plus.google.com/u/0/106313804833283549032/posts/
X1TvcxNhMWz
15. Satoshi Nakamoto: Bitcoin: A Peer-to-Peer Electronic Cash System,
At http://bitcoin.org/bitcoin.pdf
16. Satoshi Nakamoto et al.: Bitcoin QT, the original and the most
prominent bitcoin software distribution which implements a full peer-
to-peer network node. Originally developed by Satoshi Nakamoto,
core developers are Satoshi Nakamoto, Gavin Andresen, Pieter Wuille,
Nils Schneider, Jeff Garzik, Wladimir J. van der Laan and Gregory
Maxwell. Available at http://bitcoin.org/en/download with source
code at https://github.com/bitcoin/bitcoin.
17. Thomas Pornin, Deterministic Usage of the Digital Signature
Algorithm (DSA) and Elliptic Curve Digital Signature Algo-
rithm (ECDSA), IETF, August 2013, http://tools.ietf.org/html/
rfc6979
18. Dorit Ron and Adi Shamir: Quantitative Analysis of the Full Bitcoin
Transaction Graph, In Financial Cryptography and Data Security
2013, preprint available at http://eprint.iacr.org/2012/584.
19. Nils
Schneider:
Recovering
Bitcoin
private
keys
using
weak
signatures
from
the
blockchain,
Blog
entry,
28
Jan-
uary
2013,
http://www.nilsschneider.net/2013/01/28/
recovering-bitcoin-private-keys.html.
20. Technical
specification
of
the
bitcoin
protocol,
https://en.
bitcoin.it/wiki/Protocol_specification
21. Filippo Valsorda: Exploiting ECDSA Failures in the Bitcoin
Blockchain, slides presented at Hack in the Box Conference 2014,
HitbSecConf2014, Kuala Lumpur, Malaysia, on 15 October 2014,
at http://conference.hitb.org/hitbsecconf2014kul/materials/
D1T1-FilippoValsorda-
ExploitingECDSAFailuresintheBitcoinBlockchain.pdf
22. Pieter
Wuille:
the
official
specification
of
BIP032,
15
Jan
2014, https://github.com/bitcoin/bips/blob/master/bip-0032.
mediawiki
23. Pieter Wuille: presentation abut BIP032 at Bitcoin 2013 Confer-
ence San Jose, CA, May 2013, https://www.youtube.com/watch?v=
WcnMjkc31Fs
Dostları ilə paylaş: |