Twenty Years of Attacks on the RSA Cryptosystem

Dan Boneh
[email protected]

1 Introduction
The RSA cryptosystem, invented by Ron Rivest, Adi Shamir, and Len Adleman [21], was rst
publicized in the August 1977 issue of Scientic American. The cryptosystem is most commonly
used for providing privacy and ensuring authenticity of digital data. These days RSA is deployed
in many commercial systems. It is used by web servers and browsers to secure web trac, it is
used to ensure privacy and authenticity of Email, it is used to secure remote login sessions, and
it is at the heart of electronic credit-card payment systems. In short, RSA is frequently used in
applications where security of digital data is a concern.
Since its initial publication, the RSA system has been analyzed for vulnerability by many
researchers. Although twenty years of research have led to a number of fascinating attacks, none of
them is devastating. They mostly illustrate the dangers of improper use of RSA. Indeed, securely
implementing RSA is a nontrivial task. Our goal is to survey some of these attacks and describe
the underlying mathematical tools they use. Throughout the survey we follow standard naming
conventions and use Alice and Bob to denote two generic parties wishing to communicate with
each other. We use Marvin to denote a malicious attacker wishing to eavesdrop or tamper with the
communication between Alice and Bob.
We begin by describing a simplied version of RSA encryption. Let N = pq be the product of
two large primes of the same size (n=2 bits each). A typical size for N is n = 1024 bits, i.e. 309
decimal digits. Each of the factors is 512 bits. Let e; d be two integers satisfying ed = 1 mod '(N )
where '(N ) = (p , 1)(q , 1) is the order of the multiplicative group ZN . We call N the RSA
modulus, e the encryption exponent, and d the decryption exponent. The pair hN; ei is the public
key. As its name suggests, it is public and is used to encrypt messages. The pair hN; di is called
the secret key or private key and is known only to the recipient of encrypted messages. The secret
key enables decryption of ciphertexts.
A message is an integer M 2 ZN . To encrypt M , one computes C = M e mod N . To decrypt
the ciphertext, the legitimate receiver computes C d mod N . Indeed,
C d = M ed = M (mod N );
where the last equality follows by Euler's theorem1 . One denes the RSA function as x 7,!
xe mod N . If d is given, the function can be easily inverted using the above equality. We refer to d
as a trapdoor enabling one to invert the function. In this survey we study the diculty of inverting
1
Our description slightly oversimplies RSA encryption. In practice, messages are padded prior to encryption
using some randomness [1]. For instance, a simple (but insucient) padding algorithm may pad a plaintext by
M

appending a few random bits to one of the ends prior to encryption. Adding randomness to the encryption process
is necessary for proper security.

1
the RSA function without the trapdoor. We refer to this as breaking RSA. More precisely, given
the triple hN; e; C i, we ask how hard is it to compute the eth root of C modulo N = pq when the
factorization of N is unknown. Since ZN is a nite set, one may enumerate all elements of ZN until
the correct M is found. Unfortunately, this results in an algorithm with running time of order N ,
namely exponential in the size of its input, which is of the order log2 N . We are interested mostly
in algorithms with a substantially lower running time, namely on the order of nc where n = log2 N
and c is some small constant (less than 5, say). Such algorithms often perform well in practice on
the inputs in question. Throughout the paper we refer to such algorithms as ecient.
In this survey we mainly study the RSA function as opposed to the RSA cryptosystem. Loosely
speaking, the diculty of inverting the RSA function on random inputs implies that given hN; e; C i
an attacker cannot recover the plaintext M . However, a cryptosystem must resist more subtle
attacks. If hN; e; C i is given, it should be intractable to recover any information about M . This is
known as semantic security2 . We do not discuss these subtle attacks, but point out that RSA as
described above is not semantically secure: given hN; e; C i, one can easily deduce some information
about the plaintext M (for instance, the Jacobi symbol of M over N can be easily deduced from
C ). RSA can be made semantically secure by adding randomness to the encryption process [1].
The RSA function x 7,! xe mod N is an example of a trapdoor one-way function. It can be easily
computed, but (as far as we know) cannot be eciently inverted without the trapdoor d except in
special circumstances. Trapdoor one-way functions can be used for digital signatures [19]. Digital
signatures provide authenticity and nonrepudiation of electronic legal documents. For instance,
they are used for signing digital checks or electronic purchase orders. To sign a message M 2 ZN
using RSA, Alice applies her private key hN; di to M and obtains a signature S = M d mod N .
Given hM; S i, anyone can verify Alice's signature on M by checking that S e = M mod N . Since
only Alice can generate S , one may suspect that an adversary cannot forge Alice's signature.
Unfortunately, things are not so simple; extra measures are needed for proper security. Digital
signatures are an important application of RSA. Some of the attacks we survey specically target
RSA digital signatures.
An RSA key pair is generated by picking two random n2 -bit primes and multiplying them to
obtain N . Then, for a given encryption exponent e < '(N ), one computes d = e,1 mod '(N )
using the extended Euclidean algorithm. Since the set of primes is suciently dense, a random
n -bit prime can be quickly generated by repeatedly picking random n -bit integers and testing each
2 2
one for primality using a probabilistic primality test [19].

1.1 Factoring Large Integers

The rst attack on an RSA public key hN; ei to consider is factoring the modulus N . Given the
factorization of N , an attacker can easily construct '(N ), from which the decryption exponent
d = e,1 mod '(N ) can be found. We refer to factoring the modulus as a brute-force attack on
RSA. Although factoring algorithms have been steadily improving, the current state of the art is
still far from posing a threat to the security of RSA when RSA is used properly. Factoring large
integers is one of the most beautiful problems of computational mathematics [18, 20], but it is not
the topic of this article. For completeness we note that the current fastest
 factoring algorithm is
1=3 2=3
the General Number Field Sieve. Its running time on n-bit integers is exp (c + o(1))n log n
for some c < 2. Attacks on RSA that take longer than this time bound are not interesting. These
2
A source that explains semantic security and gives examples of semantically secure ciphers is [11].

2
include attacks such as exhaustive search for M and some older attacks published right after the
initial publication of RSA.
Our objective is to survey attacks on RSA that decrypt messages without directly factoring the
RSA modulus N . Nevertheless, it is worth noting that some sparse sets of RSA moduli, N = pq,
can be easily factored. For instance, if p , 1 is a product of prime factors less than B , then N can
be factored in time less than B 3 . Some implementations explicitly reject primes p for which p , 1
is a product of small primes.
As noted above, if an ecient factoring algorithm exists, then RSA is insecure. The converse is
a long standing open problem: must one factor N in order to eciently compute eth roots modulo
N ? Is breaking RSA as hard as factoring? We state the concrete open problem below.
Open Problem 1 Given integers N and e satisfying gcd(e; '(N )) = 1, dene the function fe;N :
ZN ! ZN by fe;N (x) = x1=e mod N . Is there a polynomial-time algorithm A that computes the
factorization of N given N and access to an \oracle" fe;N (x) for some e?
An oracle for f (x) evaluates the function on any input x in unit time. Recently Boneh and
Venkatesan [6] provided evidence that for small e the answer to the above problem may be \no". In
other words, for small e there may not exist a polynomial-time reduction from factoring to breaking
RSA. They do so by showing that in a certain model, a positive answer to the problem for small
e yields an ecient factoring algorithm. We note that a positive answer to Open Problem 1 gives
rise to a \chosen ciphertext attack"3 on RSA. Therefore, a negative answer may be welcome.
Next we show that exposing the private key d and factoring N are equivalent. Hence there is
no point in hiding the factorization of N from any party who knows d.
Fact 1 Let hN; ei be an RSA public key. Given the private key d, one can eciently factor the
modulus N = pq. Conversely, given the factorization of N , one can eciently recover d.
Proof A factorization of N yields '(N ). Since e is known, one can recover d. This proves the
converse statement. We now show that given d one can factor N . Given d, compute k = de , 1. By
denition of d and e we know that k is a multiple of '(N ). Since '(N ) is even, k = 2t r with r odd
and t  1. We have gk = 1 for every g 2 ZN , and therefore gk=2 is a square root of unity modulo N .
By the Chinese Remainder Theorem, 1 has four square roots modulo N = pq. Two of these square
roots are 1. The other two are x where x satises x = 1 mod p and x = ,1 mod q. Using either
one of these last two square roots, the factorization of N is revealed by computing gcd(x , 1; N ).
A straightforward argument shows that if g is chosen at random from ZN then with probability
at least 1=2 (over the choice of g) one of the elements in the sequence gk=2 ; gk=4 ; : : : ; gk=2t mod N
is a square root of unity that reveals the factorization of N . All elements in the sequence can be
eciently computed in time O(n3 ) where n = log2 N . 

2 Elementary Attacks
We begin by describing some old elementary attacks. These attacks illustrate blatant misuse of
RSA. Although many such attacks exist, we give only two examples.
3
In this context, \chosen ciphertext attack" refers to an attacker Marvin, who is given a public key h i and
N; e

access to a black box that decrypts messages of his choice. Marvin succeeds in mounting the chosen ciphertext attack
if by using the black box he can recover the private key h i.
N; d

3
2.1 Common Modulus
To avoid generating a dierent modulus N = pq for each user one may wish to x N once and
for all. The same N is used by all users. A trusted central authority could provide user i with a
unique pair ei ; di from which user i forms a public key hN; ei i and a secret key hN; di i.
At rst glance this may seem to work: a ciphertext C = M ea mod N intended for Alice cannot
be decrypted by Bob since Bob does not possess da . However, this is incorrect and the resulting
system is insecure. By Fact 1 Bob can use his own exponents eb ; db to factor the modulus N . Once
N is factored Bob can recover Alice's private key da from her public key ea . This observation, due
to Simmons, shows that an RSA modulus should never be used by more than one entity.

2.2 Blinding
Let hN; di be Bob's private key and hN; ei be his corresponding public key. Suppose an adversary
Marvin wants Bob's signature on a message M 2 ZN . Being no fool, Bob refuses to sign M . Marvin
can try the following: he picks a random r 2 ZN and sets M 0 = re M mod N . He then asks Bob
to sign the random message M 0 . Bob may be willing to provide his signature S 0 on the innocent-
looking M 0 . But recall that S 0 = (M 0 )d mod N . Marvin now simply computes S = S 0 =r mod N
and obtains Bob's signature S on the original M . Indeed,
S e = (S 0 )e =re = (M 0 )ed =re  M 0 =re = M (mod N )
This technique, called blinding, enables Marvin to obtain a valid signature on a message of his
choice by asking Bob to sign a random \blinded" message. Bob has no information as to what
message he is actually signing. Since most signature schemes apply a \one-way hash" to the message
M prior to signing [19], the attack is not a serious concern. Although we presented blinding as
an attack, it is actually a useful property of RSA needed for implementing anonymous digital cash
(cash that can be used to purchase goods, but does not reveal the identity of the person making
the purchase).

3 Low Private Exponent

To reduce decryption time (or signature-generation time), one may wish to use a small value of d
rather than a random d. Since modular exponentiation takes time linear in log2 d, a small d can
improve performance by at least a factor of 10 (for a 1024 bit modulus). Unfortunately, a clever
attack due to M. Wiener [22] shows that a small d results in a total break of the cryptosystem.
Theorem 2 (M. Wiener) Let N = pq with q < p < 2q. Let d < 31 N 1=4 . Given hN; ei with
ed = 1 mod '(N ), Marvin can eciently recover d.
Proof The proof is based on approximations using continued fractions. Since ed = 1 mod '(N ),
there exists a k such that ed , k'(N ) = 1. Therefore,
 
 e k  1
 '(N ) , d  = d'(N ) :
 

Hence, kd is an approximation of '(eN ) . Although Marvin does not know '(N ), he may use N to
p p
approximate it. Indeed, since '(N ) = N ,p,q +1 and p+q ,1 < 3 N , we have jN ,'(N )j < 3 N .
4
Using N in place of '(N ), we obtain:
   
, kd  =  ed , k'(N ) Nd
, kN + k'(N ) 
e  

N 
   p 
=  1 , k(NNd, '(N ))    3kNdN  = p
3k :
 
d N
Now, k'(N ) = ed , 1 < ed. Since e < '(N ), we see that k < d < 31 N 1=4 . Hence we obtain:
 
e
 , k  1 < 1 :

N d dN 1=4 2d2


This is a classic approximation relation. The number of fractions kd with d < N approximating Ne so
closely is bounded by log2 N . In fact, all such fractions are obtained as convergents of the continued
fraction expansion of Ne [12, Th. 177]. All one has to do is compute the log N convergents of the
continued fraction for Ne . One of these will equal kd . Since ed , k'(N ) = 1, we have gcd(k; d) = 1,
and hence kd is a reduced fraction. This is a linear-time algorithm for recovering the secret key d.

Since typically N is 1024 bits, it follows that d must be at least 256 bits long in order to avoid
this attack. This is unfortunate for low-power devices such as \smartcards", where a small d would
result in big savings. All is not lost however. Wiener describes a number of techniques that enable
fast decryption and are not susceptible to his attack:
Large e: Suppose instead of reducing e modulo '(N ), one uses hN; e0 i for the public key, where
e0 = e + t
'(N ) for some large t. Clearly e0 can be used in place of e for message encryption.
However, when a large value of e is used, the k in the above proof is no longer small. A simple
calculation shows that if e0 > N 1:5 then no matter how small d is, the above attack cannot
be mounted. Unfortunately, large values of e result in increased encryption time.
Using CRT: An alternate approach is to use the Chinese Remainder Theorem (CRT). Suppose
one chooses d such that both dp = d mod (p , 1) and dq = d mod (q , 1) are small, say
128 bits each. Then fast decryption of a ciphertext C can be carried out as follows: rst
compute Mp = C dp mod p and Mq = C dq mod q. Then use the CRT to compute the unique
value M 2 ZN satisfying M = Mp mod p and M = Mq mod q. The resulting M satises
M = C d mod N as required. The point is that although dp and dq are small, the value of
d mod '(N ) can be large, i.e., on the order of '(N ). As a result, the attack of Theorem 2
does not apply. We note
, that pif hN;
p ei
is given, there exists an attack enabling an adversary
to factor N in time O min( dp ; dq ) . Hence, dp and dq cannot be made too small.
We do not know whether either of these methods is secure. All we know is that Wiener's attack
is ineective against them. Theorem 2 was recently improved by Boneh and Durfee [4], who show
that as long as d < N 0:292 , an adversary can eciently recover d from hN; ei. These results show
that Wiener's bound is not tight. It is likely that the correct bound is d < N 0:5 . At the time of
this writing, this is an open problem.
Open Problem 2 Let N = pq and d < N 0:5. If Marvin is given hN; ei with ed = 1 mod '(N )
and e < '(N ), can he eciently recover d?

5
4 Low Public Exponent
To reduce encryption or signature-verication time, it is customary to use a small public exponent
e. The smallest possible value for e is 3, but to defeat certain attacks the value e = 216 + 1 = 65537
is recommended. When the value 216 + 1 is used, signature verication requires 17 multiplications,
as opposed to roughly 1000 when a random e  '(N ) is used. Unlike the attack of the previous
section, attacks that apply when a small e is used are far from a total break.

4.1 Coppersmith's Theorem

The most powerful attacks on low public exponent RSA are based on a theorem due to Copper-
smith [7]. Coppersmith's theorem has many applications, only some of which will be covered here.
The proof uses the LLL lattice basis reduction algorithm [17] as explained below.
Theorem 3 (Coppersmith)
1 ,
Let N be an integer and f 2 Z[x] be a monic polynomial of degree
d. Set X = N for some   0. Then, given hN; f i Marvin can eciently nd all integers
d
jx0 j < X satisfying f (x0) = 0 mod N . The running time is dominated by the time it takes to run
the LLL algorithm on a lattice of dimension O(w) with w = min(1=; log2 N ).
The theorem provides an algorithm for eciently nding all roots of f modulo N that are
less than X = N 1=d . As X gets smaller, the algorithm's running time decreases. The theorem's
strength is its ability to nd small roots of polynomials modulo a composite N . When working
modulo a prime, there is no reason to use Coppersmith's theorem since other, far better, root-nding
algorithms exist.
We sketch the main ideas behind the proof of Coppersmith's theorem. P
We follow a simplied
approach due to Howgrave-Graham [14]. Given a polynomial h ( x ) = ai xi 2 Z[x], dene khk2 =
P
i jai j . The proof relies on the following observation.
2

Lemma 4 Letph(x) 2 Z[x] be a polynomial of degree d and let X be a positive integer. Suppose
kh(xX )k < N= d. If jx0j < X satises h(x0 ) = 0 mod N , then h(x0 ) = 0 holds over the integers.
Proof Observe from the Schwarz inequality that
X  X  x i  X   x i 
jh(x )j =
0


i   i
ai x0  =  aiX X  0   i
ai X X 
 0 
X  p
 ai X i  k
d h(xX ) < N:k
Since h(x0 ) = 0 mod N , we conclude that h(x0 ) = 0. 
The lemma states that if h is a polynomial with low norm, then all small roots of h mod N are
also roots of h over the integers. The lemma suggests that to nd a small root x0 of f (x) mod N we
should look for another polynomial h 2 Z[x] with small norm having the same roots as f modulo
N . Then x0 will be a root of h over the integers and can be easily found. To do so, we may search
for a polynomial g 2 Z[x] such that h = gf has low norm, i.e., norm less than N . This amounts to
searching for an integer linear combination of the polynomials f; xf; x2 f; : : : ; xr f with low norm.
Unfortunately, most often there is no nontrivial linear combination with suciently small norm.

6
 Coppersmith found a trick to solve the problem: if f (x0 ) = 0 mod N , then f (x0 )k = 0 mod N k
for any k. More generally, dene the following polynomials:
gu;v (x) = N m,v xu f (x)v
for some predened m. Then x0 is a root of gu;v (x) modulo N m for any u  0 and 0  v  m.
To use Lemma 4 we must nd an integer linear combination h(x) of the polynomials gu;v (x) such
that h(xX ) has norm less than N m (recall that X is an upper bound on x0 satisfying X  N 1=d ).
Thanks to the relaxed upper bound on the norm (N m rather than N ), one can show that for
suciently large m, there always exists a linear combination h(x) satisfying the required bound.
Once h(x) is found, Lemma 4 implies that it has x0 as a root over the integers. Consequently x0
can be easily found.
It remains to show how to nd h(x) eciently. To do so, we must rst state a few basic facts
about lattices in Zw. We refer to [17] for a concise introduction to the topic. Let u1 ; : : : ; uw 2 Zw
be linearly independent vectors. A (full-rank) lattice L spanned by hu1 ; : : : ; uw i is the set of all
integer linear combinations of u1 ; : : : ; uw . The determinant of L is dened as the determinant of
the w  w square matrix whose rows are the vectors u1 ; : : : ; uw .
In our case, we view the polynomials gu;v (xX ) as vectors and study the lattice L spanned by
them. We let v = 0; : : : ; m and u = 0; : : : ; d , 1, and hence the lattice has dimension w = d(m + 1).
For example, when f is a quadratic monic polynomial and m = 3, the resulting lattice is spanned
by the rows of the following matrix:
1 x x2 x3 x4 x5 x6 x7
g0;0 (xX ) 2 N 3 3

g1;0 (xX ) 66 XN 3 7
7
g0;1 (xX ) 6 6  XN 2 2 7
7
g1;1 (xX ) 66   XN 3 2 7
7
g0;2 (xX ) 6 6    XN 4 7
7
g1;2 (xX ) 6 6     XN 5 7
7
g0;3 (xX ) 4       X 6 5
g1;3 (xX )       X 7

The entries  correspond to coecients of the polynomials whose value we ignore. All empty
entries are zero. Since the matrix is triangular, its determinant is the product of the elements on
the diagonal (which are explicitly given above). Our objective is to nd short vectors in this lattice.
A classic result of Hermite states that any lattice L of dimension w contains a nonzero point
v 2 L whose L2 norm satises kvk  w det(L)1=w , where w is a constant depending only on w.
Hermite's bound can be used to show that for large enough m our lattice contains vectors of norm
less than N m , as required. The question is whether we can eciently construct a short vector in
L whose length is not much larger than the Hermite bound. The LLL algorithm is an ecient
algorithm that does precisely that.
Fact 5 (LLL) Let L be a lattice spanned by hu1 ; : : : ; uw i. When hu1 ; : : : ; uw i are given as input,
then the LLL algorithm outputs a point v 2 L satisfying
kvk  2w=4 det(L)1=w :
The running time for LLL is quartic in the length of the input.

7
 The LLL algorithm (named after its inventors L. Lovasz, A. Lenstra, and H. Lenstra Jr.) has
many applications in both computational number theory and cryptography. Its discovery in 1982
provided an ecient algorithm for factoring polynomials over the integers and, more generally,
over number rings. LLL is frequently used to attack various cryptosystems. For instance, many
cryptosystems based on the \knapsack problem" have been broken using LLL.
Using LLL, we can complete the proof of Coppersmith's theorem. To ensure that the vector
produced by LLL satises the bound of Lemma 4 we need
p
2w=4 det(L)1=w < N m = w;
where w = d(m +1) is the dimension of L1 . A routine calculation shows that for large enough m the
bound is satised. Indeed, when X = N d , , it suces to take m = O(k=d) with k = min( 1 ; log N ).
Consequently, the running time is dominated by running LLL on a lattice of dimension O(k), as
required.
A natural question is whether Coppersmith's theorem can be applied to bivariate and multi-
variate polynomials. If f (x; y) 2 ZN [x; y] is given for which there exists a root (x0 ; y0 ) with jx0 y0 j
suitably bounded, can Marvin eciently nd (x0 ; y0 )? Although the same technique appears to
work for some bivariate polynomials, it is currently an open problem to prove it. As an increasing
number of results depend on a bivariate extension of Coppersmith's theorem, a rigorous algorithm
will be very useful.
Open Problem 3 Find general conditions under which Coppersmith's theorem can be generalized
to bivariate polynomials.

4.2 Hastad's Broadcast Attack

As a rst application of Coppersmith's theorem, we present an improvement to an old attack due
to Hastad [13]. Suppose Bob wishes to send an encrypted message M to a number of parties
P1 ; P2 ; : : : ; Pk . Each party has its own RSA key hNi ; ei i. We assume M is less than all the Ni 's.
Naively, to send M , Bob encrypts it using each of the public keys and sends out the ith ciphertext
to Pi . An attacker Marvin can eavesdrop on the connection out of Bob's sight and collect the k
transmitted ciphertexts.
For simplicity, suppose all public exponents ei are equal to 3. A simple argument shows that
Marvin can recover M if k  3. Indeed, Marvin obtains C1 ; C2 ; C3 , where
C1 = M 3 mod N1 ; C2 = M 3 mod N2 ; C3 = M 3 mod N3 :
We may assume that gcd(Ni ; Nj ) = 1 for all i 6= j since otherwise Marvin can factor some of the
Ni 's. Hence, applying the Chinese Remainder Theorem (CRT) to C1 ; C2 ; C3 gives a C 0 2 ZN1N2N3
satisfying C 0 = M 3 mod N1 N2 N3 . Since M is less than all the Ni 's, we have M 3 < N1 N2 N3 . Then
C 0 = M 3 holds over the integers. Thus, Marvin may recover M by computing the real cube root of
C 0 . More generally, if all public exponents are equal to e, Marvin can recover M as soon as k  e.
The attack is feasible only when a small e is used.
Hastad [13] describes a far stronger attack. To motivate Hastad's result, consider a naive defense
against the above attack. Rather than broadcasting the encryption of M , Bob could \pad" the
message prior to encryption. For instance, if M is m bits long, Bob could send Mi = i2m + M
to party Pi . Since Marvin obtains encryptions of dierent messages, he cannot mount the attack.
8
Unfortunately, Hastad showed that this linear padding is insecure. In fact, he proved that applying
any xed polynomial to the message prior to encryption does not prevent the attack.
Suppose that for each of the participants P1 ; : : : ; Pk , Bob has a xed public polynomial fi 2
ZNi [x]. To broadcast a message M , Bob sends the encryption of fi (M ) to party Pi . By eavesdrop-
ping, Marvin learns Ci = fi (M )ei mod Ni for i = 1; : : : ; k. Hastad showed that if enough parties
are involved, Marvin can recover the plaintext M from all the ciphertexts. The following theorem
is a stronger version of Hastad's original result.
Theorem 6 (Hastad) Let N1; : : : Nk be pairwise relatively prime integers and set Nmin = mini (Ni).
Let gi 2 ZNi [x] be k polynomials of maximum degree d. Suppose there exists a unique M < Nmin
satisfying
gi (M ) = 0 mod Ni for all i = 1; : : : ; k:
Under the assumption that k > d, one can eciently nd M given hNi ; gi iki=1 .
Proof Let N = N1


Nk . We may assume that all gi 's are monic. (Indeed if, for some i,
the leading coecient of gi is not invertible in ZNi , then the factorization of Ni is exposed.) By
multiplying each gi by the appropriate power of x, we may assume they all have degree d. Construct
the polynomial
k
g(x) = Tigi (x); where Ti = 10 mod
X Nj if i = j
i=1
mod Nj if i 6= j:
The Ti 's are integers known as the Chinese Remainder Coecients. Then g(x) must be monic
since it is monic modulo all the Ni . Its degree is d. Furthermore, we know that g(M ) = 0 mod N .
Theorem 6 now follows from Theorem 3 since M < Nmin  N 1=k < N 1=d . 
The theorem shows that a system of univariate equations modulo relatively prime composites
can be eciently solved, assuming suciently many equations are provided. By setting gi =
fiei , Ci mod Ni , we see that Marvin can recover M from the given ciphertexts whenever the
number of parties is at least d, where d is the maximum of ei deg(fi ) over all i = 1; : : : ; k. In
particular, if all ei 's are equal to e and Bob sends out linearly related messages, then Marvin can
recover the plaintext as soon as k > e.
Hastad's original theorem is weaker than the one stated above. Rather than d polynomials,
Hastad required d(d + 1)=2 polynomials. Hastad's proof is similar to the proof of Coppersmith's
theorem described in the previous section. However, Hastad does not use powers of g in the lattice
and consequently obtains a weaker bound.
To conclude this section we note that to properly defend against the broadcast attack above,
one must use a randomized pad [1] rather than a xed one.

4.3 Franklin-Reiter Related Message Attack

Franklin and Reiter [8] found a clever attack when Bob sends Alice related encrypted messages
using the same modulus. Let hN; ei be Alice's public key. Suppose M1 ; M2 2 ZN are two distinct
messages satisfying M1 = f (M2 ) mod N for some publicly known polynomial f 2 ZN [x]. To send
M1 and M2 to Alice, Bob may naively encrypt the messages and transmit the resulting ciphertexts
C1 ; C2 . We show that given C1 ; C2 , Marvin can easily recover M1 ; M2 . Although the attack works
for any small e, we state the following lemma for e = 3 in order to simplify the proof.

9
Lemma 7 (FR) Set e = 3 and let hN; ei be an RSA public key. Let M 6= M 2 ZN satisfy 1 2
M = f (M ) mod N for some linear polynomial f = ax + b 2 ZN [x] with b 6= 0. Then, given
1 2
hN; e; C ; C ; f i, Marvin can recover M ; M in time quadratic in log N .
1 2 1 2

Proof To keep this part of the proof general, we state it using an arbitrary e (rather than
restricting to e = 3). Since C1 = M1e mod N , we know that M2 is a root of the polynomial
g1 (x) = f (x)e , C1 2 ZN [x]. Similarly, M2 is a root of g2(x) = xe , C2 2 ZN [x]. The linear factor
x , M2 divides both polynomials. Therefore, Marvin may use the Euclidean algorithm4 to compute
the gcd of g1 and g2 . If the gcd turns out to be linear, M2 is found. The gcd can be computed in
quadratic time in e and log N .
We show that when e = 3 the gcd must be linear. The polynomial x3 , C2 factors modulo
both p and q into a linear factor and an irreducible quadratic factor (recall that gcd(e; '(N )) = 1
and hence x3 , C2 has only one root in ZN ). Since g2 cannot divide g1 , the gcd must be linear.
For e > 3 the gcd is almost always linear. However, for some rare M1 ; M2 , and f , it is possible to
obtain a nonlinear gcd, in which case the attack fails. 
For e > 3 the attack takes time quadratic in e. Consequently, it can be applied only when a
small public exponent e is used. For large e the work in computing the gcd is prohibitive. It is
an interesting question (though likely to be dicult) to devise such an attack for arbitrary e. In
particular, can the gcd of g1 and g2 above be found in time polynomial in log e?

4.4 Coppersmith's Short Pad Attack

The Franklin-Reiter attack might seem a bit articial. After all, why should Bob send Alice the
encryption of related messages? Coppersmith strengthened the attack and proved an important
result on padding [7].
A naive random padding algorithm might pad a plaintext M by appending a few random bits to
one of the ends. The following attack points out the danger of such simplistic padding. Suppose Bob
sends a properly-padded encryption of M to Alice. An attacker, Marvin, intercepts the ciphertext
and prevents it from reaching its destination. Bob notices that Alice did not respond to his message
and decides to resend M to Alice. He randomly pads M and transmits the resulting ciphertext.
Marvin now has two ciphertexts corresponding to two encryptions of the same message using two
dierent random pads. The following theorem shows that although he does not know the pads
used, Marvin is able to recover the plaintext.
Theorem 8 Let hN; ei be a public RSA key where N is n-bits long. Set m = bn=e2 c. Let M 2 ZN
be a message of length at most n , m bits. Dene M1 = 2m M + r1 and M2 = 2m M + r2 , where
r1 and r2 are distinct integers with 0  r1 ; r2 < 2m . If Marvin is given hN; ei and the encryptions
C1 ; C2 of M1 ; M2 (but is not given r1 or r2 ), he can eciently recover M .
Proof Dene g1 (x; y) = xe , C1 and g2 (x; y) = (x + y)e , C2. We know that when y = r2 , r1 ,
these polynomials have M1 as a common root. In other words,
= r2 , r1 is a root of the \resul-
tant" h(y) = resx (g1 ; g2 ) 2 ZN [y]. The degree of h is at most e2 . Furthermore, j
j < 2m < N 1=e2 .
Hence,
is a small root of h modulo N , and Marvin can eciently nd it using Coppersmith's
theorem (Theorem 3). Once
is known, the Franklin-Reiter attack of the previous section can be
4
Although ZN [ ] is not a Euclidean ring, the standard Euclidean algorithm can still be applied to polynomials in
x

ZN [ ]. One can show that if the algorithm \breaks" in any way, then the factorization of is exposed.
x N

10
used to recover M2 and consequently M . 
When e = 3 the attack can be mounted as long as the pad length is less than 1=9th the message
length. This is an important result. Note that for the recommended value of e = 65537, the attack
is useless against standard moduli sizes.

4.5 Partial Key Exposure Attack

Let hN; di be a private RSA key. Suppose by some means Marvin is able to expose a fraction of
the bits of d, say a quarter of them. Can he reconstruct the rest of d? Surprisingly, the answer
is positive when the corresponding
p public key is small. Recently Boneh, Durfee, and Frankel [5]
showed that as long as e < N , it is possible to reconstruct all of d from just a fraction of its bits.
These results illustrate the importance of safeguarding the entire private RSA key.
Theorem 9 (BDF) Let hN; di be a private RSA key in which N is n bits long. Given the dn=4e
least signicant bits of d, Marvin can reconstruct all of d in time linear in e log2 e.
The proof relies on yet another beautiful theorem due to Coppersmith [7].
Theorem 10 (Coppersmith) Let N = pq be an n-bit RSA modulus. Then given the n=4 least
signicant bits of p or the n=4 most signicant bits of p, one can eciently factor N .
Theorem 9 readily follows from Theorem 10. In fact, by denition of e and d, there exists an
integer k such that
ed , k(N , p , q + 1) = 1:
Since d < '(N ), we must have 0 < k  e. Reducing the equation modulo 2n=4 and setting q = N=p,
we obtain
(ed)p , kp(N , p + 1) + kN = p (mod 2n=4 ):
Since Marvin is given the n=4 least signicant bits of d, he knows the value of ed mod 2n=4 . Con-
sequently, he obtains an equation in k and p. For each of the e possible values of k, Marvin solves
the quadratic equation in p and obtains a number of candidate values for p mod 2n=4 . For each
of these candidate values, he runs the algorithm of Theorem 10 to attempt to factor N . One can
show that the total number of candidate values for p mod 2n=4 is at most e log2 e. Hence after at
most e log2 e attempts, N will be factored.
Theorem 9 is p known as a partial key-exposure attack. Similar attacks exist for larger values of
e as long as e < N . However the techniques are a bit more complex [5]. It is interesting that
discrete log-based cryptosystems, such as the ElGamal public key system, do not seem susceptible
to partial key exposure. Indeed, if gx mod p and a constant fraction of the bits of x are given, there
is no known polynomial-time algorithm to compute the rest of x.
To conclude the section we show that when the encryption exponent e is small, the RSA system
leaks half the most signicant bits of the corresponding private key d. To see this, consider once
again the equation ed , k(N , p , q + 1) = 1 for an integer 0 < k  e. Given k, Marvin may easily
compute
db = b(kN + 1)=ec:
Then p p
jdb, dj  k(p + q)=e  3k N=e < 3 N:

11
Hence, db is a good approximation for d. The bound shows that, for most d, half the most signicant
bits of db are equal to those of d. Since there are only e possible values for k, Marvin can construct
a small set of size e such that one of the elements in the set is equal to half the most signicant
bits of d. The case e = 3 is especially interesting. In this case one can show that always k = 2 and
hence the system completely leaks half the most signicant bits of d.

5 Implementation Attacks
We turn our attention to an entirely dierent class of attacks. Rather than attacking the underlying
structure of the RSA function, these attacks focus on the implementation of RSA.

5.1 Timing Attacks

Consider a smartcard that stores a private RSA key. Since the card is tamper resistant, an attacker
Marvin may not be able to examine its contents and expose the key. However, a clever attack due
to Kocher [16] shows that by precisely measuring the time it takes the smartcard to perform an
RSA decryption (or signature), Marvin can quickly discover the private decryption exponent d.
We explain how to mount the attack against a simple implementation of RSA using the
\repeated
P
squaring algorithm". Let d = dn dn,1 : : : d0 be the binary representation of d (i.e.,
d = ni=0 2i di with di 2 f0; 1g). The repeated squaring algorithm computes C =QM d mod N , using
at most 2n modular multiplications. It is based on the observation that C = ni=0 M 2i di mod N .
The algorithm works as follows:
Set z equal to M and C equal to 1. For i = 0; : : : ; n, do these steps:
(1) if di = 1 set C equal to C
z mod N ,
(2) set z equal to z 2 mod N .
At the end, C has the value M d mod N .
The variable z runs through the set of values M 2i mod N for i = 0; : : : ; n. The variable C \collects"
the appropriate powers in the set to obtain M d mod N .
To mount the attack, Marvin asks the smartcard to generate signatures on a large number of
random messages M1 ; : : : ; Mk 2 ZN and measures the time Ti it takes the card to generate each of
the signatures.
The attack recovers bits of d one at a time beginning with the least signicant bit. We know d is
odd. Thus d0 = 1. Consider the second iteration. Initially z = M 2 mod N and C = M . If d1 = 1,
the smartcard computes the product C
z = M
M 2 mod N . Otherwise, it does not. Let ti be the
time it takes the smartcard to compute Mi
Mi2 mod N . The ti 's dier from each other since the
time to compute Mi
Mi2 mod N depends on the value of Mi (simple modular reduction algorithms
take a dierent amount of time depending on the value being reduced). Marvin measures the ti 's
oine (prior to mounting the attack) once he obtains the physical specications of the card.
Kocher observed that when d1 = 1, the two ensembles fti g and fTi g are correlated. For instance,
if, for some i; ti is much larger than its expectation, then Ti is also likely to be larger than its
expectation. On the other hand, if d1 = 0, the two ensembles fti g and fTi g behave as independent
random variables. By measuring the correlation, Marvin can determine whether d1 is 0 or 1.

12
Continuing in this way, he can recover d2 , d3 , and so on. Note that when a low public exponent e
is used, the partial key exposure attack of the previous section shows that Kocher's timing attack
need only be employed until a quarter of the bits of d are discovered.
There are two ways to defend against the attack. The simplest is to add appropriate delay so
that modular exponentiation always takes a xed amount of time. The second approach, due to
Rivest, is based on blinding. Prior to decryption of M the smartcard picks a random r 2 ZN and
computes M 0 = M
re mod N . It then applies d to M 0 and obtains C 0 = (M 0 )d mod N . Finally,
the smartcard sets C = C 0 =r mod N . With this approach, the smartcard is applying d to a random
message M 0 unknown to Marvin. As a result, Marvin cannot mount the attack.
Kocher recently discovered another attack along these lines called power cryptanalysis. Kocher
showed that by precisely measuring the smartcard's power consumption during signature gener-
ation, Marvin can often easily discover the secret key. As it turns out, during a multi-precision
multiplication the card's power consumption is higher than normal. By measuring the length of
high consumption periods, Marvin can easily determine if in a given iteration the card performs
one or two multiplications, thus exposing the bits of d.

5.2 Random Faults

Implementations of RSA decryption and signatures frequently use the Chinese Remainder Theorem
to speed up the computation of M d mod N . Instead of working modulo N , the signer Bob rst
computes the signatures modulo p and q and then combines the results using the Chinese Remainder
Theorem. More precisely, Bob rst computes
Cp = M dp mod p and Cq = M dq mod q;
where dp = d mod (p , 1) and dq = d mod (q , 1). He then obtains the signature C by setting
C = T1 Cp + T2Cq (mod N );
where    
T1 = 10 mod p and T 0 mod p :
2 =
mod q 1 mod q
The running time of the last CRT step is negligible compared to the two exponentiations. Note that
p and q are half the length of N . Since simple implementations of multiplication take quadratic time,
multiplication modulo p is four times faster than modulo N . Furthermore, dp is half the length
of d and consequently computing M dp mod p is eight times faster than computing M d mod N .
Overall signature time is thus reduced by a factor of four. Many implementations use this method
to improve performance.
Boneh, DeMillo, and Lipton [3] observed that there is an inherent danger in using the CRT
method. Suppose that while generating a signature, a glitch on Bob's computer causes it to
miscalculate in a single instruction. Say, while copying a register from one location to another, one
of the bits is ipped. (A glitch may be caused by ambient electromagnetic interference or perhaps
by a rare hardware bug, like the one found in an early version of the Pentium chip.) Given an
invalid signature, an adversary Marvin can easily factor Bob's modulus N .
We present a version of the attack as described by A. K. Lenstra. Suppose a single error
occurs while Bob is generating a signature. As a result, exactly one of Cp or Cq will be computed
13
incorrectly. Say Cp is correct, but Cbq is not. The resulting signature is Cb = T1 Cp + T2 Cbq . Once
Marvin receives Cb, he knows it is a false signature since Cbe 6= M mod N . However, notice that
Cbe = M mod p while Cbe 6= M mod q:
As a result, gcd(N; Cbe , M ) exposes a nontrivial factor of N .
For the attack to work, Marvin must have full knowledge of M . Namely, we are assuming
Bob does not use any random padding procedure. Random padding prior to signing defeats the
attack. A simpler defense is for Bob to check the generated signature before sending it out to the
world. Checking is especially important when using the CRT speedup method. Random faults are
hazardous to many cryptographic systems. Many systems, including a non-CRT implementation
of RSA, can be attacked using random faults [3]. However, these results are far more theoretical.

5.3 Bleichenbacher's Attack on PKCS 1

Let N be an n-bit RSA modulus and M be an m bit message with m < n. Before applying RSA
encryption it is natural to pad the message M to n bits by appending random bits to it. An old
version of a standard known as Public Key Cryptography Standard 1 (PKCS 1) uses this approach.
After padding, the message looks as follows:
02 Random 00 M
The resulting message is n bits long and is directly encrypted using RSA. The initial block con-
taining \02" is 16 bits long and is there to indicate that a random pad has been added to the
message.
When a PKCS 1 message is received by Bob's machine, an application (e.g., a web browser)
decrypts it, checks the initial block, and strips o the random pad. However, some applications
check for the \02" initial block and if it is not present they send back an error message saying
\invalid ciphertext". Bleichenbacher [2] showed that this error message can lead to disastrous
consequences: using the error message, an attacker Marvin can decrypt ciphertexts of his choice.
Suppose Marvin intercepts a ciphertext C intended for Bob and wants to decrypt it. To mount
the attack, Marvin picks a random r 2 ZN , computes C 0 = rC mod N , and sends C 0 to Bob's
machine. An application running on Bob's machine receives C 0 and attempts to decrypt it. It
either responds with an error message or does not respond at all (if C 0 happens to be properly
formatted). Hence, Marvin learns whether the most signicant 16 bits of the decryption of C 0 are
equal to 02. In eect, Marvin has an oracle that tests for him whether the 16 most signicant bits
of the decryption of rC mod N are equal to 02, for any r of his choice. Bleichenbacher showed that
such an oracle is sucient for decrypting C .

6 Conclusions
Two decades of research into inverting the RSA function produced some insightful attacks, but no
devastating attack has ever been found. The attacks discovered so far mainly illustrate the pitfalls
to be avoided when implementing RSA. At the moment it appears that proper implementations
can be trusted to provide security in the digital world.

14
 We categorized attacks on RSA into four categories: (1) elementary attacks that exploit blatant
misuse of the system, (2) low private exponent attacks serious enough that a low private exponent
should never be used, (3) low public exponent attacks, (4) and attacks on the implementation.
These last attacks illustrate that a study of the underlying mathematical structure is insucient.
Desmedt and Odlyzko [10], Joye and Quisquater [15] and deJonge and Chaum [9] describe some
additional attacks. Throughout the paper we observed that many attacks can be defeated by
properly padding the message prior to encryption or signing.

Acknowledgments
I thank Susan Landau for encouraging me to write the survey and Tony Knapp for his help in
editing the manuscript. I am also grateful to Mihir Bellare, Igor Shparlinski, and R. Venkatesan
for comments on an earlier draft.

References
[1] M. Bellare and P. Rogaway. Optimal asymmetric encryption. In EUROCRYPT '94, volume
950 of Lecture Notes in Computer Science, pages 92{111. Springer-Verlag, 1994.
[2] D. Bleichenbacher. Chosen ciphertext attacks against protocols based on the RSA encryption
standard PKCS #1. In CRYPTO '98, volume 1462 of Lecture Notes in Computer Science,
pages 1{12. Springer-Verlag, 1998.
[3] D. Boneh, R. DeMillo, and R. Lipton. On the importance of checking cryptographic protocols
for faults. In EUROCRYPT '97, volume 1233 of Lecture Notes in Computer Science, pages
37{51. Springer-Verlag, 1997.
[4] D. Boneh and G. Durfee. New results on cryptanalysis of low private exponent RSA. Preprint,
1998.
[5] D. Boneh, G. Durfee, and Y. Frankel. An attack on RSA given a fraction of the private
key bits. In AsiaCrypt '98, volume 1514 of Lecture Notes in Computer Science, pages 25{34.
Springer-Verlag, 1998.
[6] D. Boneh and R. Venkatesan. Breaking RSA may not be equivalent to factoring. In EURO-
CRYPT '98, volume 1403 of Lecture Notes in Computer Science, pages 59{71. Springer-Verlag,
1998.
[7] D. Coppersmith. Small solutions to polynomial equations, and low exponent RSA vulnerabil-
ities. Journal of Cryptology, 10:233{260, 1997.
[8] D. Coppersmith, M. Franklin, J. Patarin, and M. Reiter. Low-exponent RSA with related
messages. In EUROCRYPT '96, volume 1070 of Lecture Notes in Computer Science, pages
1{9. Springer-Verlag, 1996.
[9] W. de Jonge and D. Chaum. Attacks on some RSA signatures. In Crypto '85, volume 218 of
Lecture Notes in Computer Science, pages 18{27. Springer-Verlag, 1986.

15
[10] Y. Desmedt and A. Odlyzko. A chosen text attack on the rsa cryptosystem and some discrete
logarithm schemes. In CRYPTO '85, Lecture Notes in Computer Science, pages 516{522.
Springer-Verlag, 1985.
[11] S. Goldwasser. The search for provably secure cryptosystems. In Cryptology and computational
number theory, volume 42 of Proceedings of the 42nd Symposium in Applied Mathematics.
American Mathematical Society, 1990.
[12] G. H. Hardy and E. M. Wright. An Introduction to the Theory of Numbers. Oxford Clarendon
Press, 1975. fourth edition.
[13] J. Hastad. Solving simultaneous modular equations of low degree. SIAM J. of Computing,
17:336{341, 1988.
[14] N. Howgrave-Graham. Finding small roots of univariate modular equations revisited. In
Cryptography and Coding, volume 1355 of Lecture Notes in Computer Science, pages 131{142.
Springer-Verlag, 1997.
[15] M. Joye and J.-J. Quisquater. On the importance of securing your bins: The garbage-man-
in-the-middle attack. In 4th ACM Conference on Computer and Communications Security,
pages 135{141. ACM Press, 1997.
[16] P. Kocher. Timing attacks on implementations of Die-Hellman, RSA, DSS, and other sys-
tems. In CRYPTO '96, volume 1109 of Lecture Notes in Computer Science, pages 104{113.
Springer-Verlag, 1996.
[17] L. Lovasz. An Algorithmic Theory of Number, Graphs and Convexity. SIAM Publications,
1986.
[18] A. K. Lenstra and H. W. Lenstra, Jr. Algorithms in number theory. In Handbook of Theoret-
ical Computer Science (Volume A: Algorithms and Complexity), chapter 12, pages 673{715.
Elsevier and MIT Press, 1990.
[19] A. Menezes, P. van Oorschot, and S. Vanstone. Handbook of Applied Cryptography. CRC,
1996.
[20] C. Pomerance. A tale of two sieves. Notices Amer. Math. Soc., 43:1473{1485, 1996.
[21] R. L. Rivest, A. Shamir, and L. Adleman. A method for obtaining digital signatures and
public key cryptosystems. Commun. of the ACM, 21:120{126, 1978.
[22] M. Wiener. Cryptanalysis of short RSA secret exponents. IEEE Transactions on Information
Theory, 36:553{558, 1990.

16