Method for elliptic curve point multiplication

Abstract

The method comprises three stages. In the first stage, randomly selected point representations are stored in variables. In the second stage, a right-to-left loop is executed that modifies the variable values in dependency of a multiplier. In the last stage, the result is calculated from the modified variable values.

Description

TECHNICAL FIELD OF THE INVENTION

The invention describes an elliptic curve point multiplication method with resistance against side-channel attacks, which are a big threat for use in cryptography, e.g. for key exchange, encryption, or for digital signatures.

PRIOR ART

Implementations of elliptic curve cryptosystems may be vulnerable to side-channel attacks ([1],[2]) where adversaries can use power consumption measurements or similar observations to derive information on secret scalars e in point multiplications eP.

One distinguishes between differential side-channel attacks, which require correlated measurements from multiple point multiplications, and simple side-channel attacks, which directly interpret data obtained during a single point multiplication. Randomisation can be used as a countermeasure against differential side-channel attacks.

In particular, for elliptic curve cryptography, projective randomisation is a simple and effective tool ([3]):

If (X,Y,Z) represents the point whose affine coordinates are (X/Z², Y/Z³) another representation of the same point that cannot be predicted by the adversary is obtained by substituting (r²X, r³Y, rZ) with a randomly chosen secret non-zero field element r. (When starting from an affine representation (X,Y), this simplifies to (r²X, r³Y, r).)

Simple side-channel attacks can be easily performed because usually the attacker can tell apart point doublings from general point additions.

Thus point multiplication should be implemented using a fixed sequence of point operations that does not depend on the particular scalar.

Note that it is reasonable to assume that point addition and point subtraction are uniform to the attacker as point inversion is nearly immediate (dummy inversions can be inserted to obtain the same sequence of operations for point additions as for point subtractions).

Various point multiplication methods have been proposed that use an alternating sequence of doublings and additions:

The simplest approach uses a binary point multiplication method with dummy additions inserted to avoid dependencies on scalar bits ([3]); however as noted in [4] it may be easy for adversaries to determine which additions are dummy operations, so it is not clear that this method provides sufficient security. For odd scalars, a variant of binary point multiplication can be used where the scalar is represented in balanced binary representation (digits −1 and +1) ([5]). Also Montgomery's binary point multiplication method ([6]), which maintains an invariant Q₁−Q₀=P while computing eP using two variables Q₀, Q₁, can be adapted for implementing point multiplication with a fixed sequence of point operations ([7], [8], [9], [10], [11]).

With this approach, specific techniques can be used to speed up point arithmetic:

The doubling and addition steps can be combined; y-coordinates of points may be omitted during the computation ([6],[9],[10],[11]); and on suitable hardware, parallel execution can be conveniently used for improved efficiency ([10],[11]).

All of the above point multiplication methods are binary. Given sufficient memory, efficiency can be improved by using 2^w-ary point multiplication methods. Here, the scalar e is represented in base 2^w using digits b_i from some digit set B:

$e=\sum _{0\le i\le l}^{}{b}_i{2}^{\mathrm{wi}}$

A simple way to obtain a uniform sequence of doublings and additions (namely, one addition after w doublings in the main loop of the point multiplication algorithm) is to use 2^w-ary point multiplication as usual (first compute and store bP for each bεB, then compute eP using this precomputed table), but to insert a dummy addition whenever a zero digit is encountered.

However, as noted above for the binary case, the dummy addition approach may not be secure.

This problem can be avoided (given w≧2) by using a representation of e without digit value 0, such asB={−2^w,1,2, . . ., 2^w−1}as proposed in [4], orB={−2^w, ±1,±2, . . . ,±(2^w−2),2^w−1}for improved efficiency as proposed in [12].

A remaining problem in the method of [4] and [12] is that the use of a fixed table may allow for statistical attacks: If the same point from the table is used in a point addition whenever the same digit value occurs, this may help adversaries to find out which of the digits b, have the same value (cf. the attacks on modular exponentiation using fixed tables in [13] and [14]).

This problem can be countered by performing, whenever the table is accessed, a projective randomisation of the table value that has been used.

This will avoid a fixed table, but at the price of reduced efficiency.

BACKGROUND OF THE INVENTION

This invention is a variant of 2^w-ary point multiplication with resistance against side-channel attacks that avoids a fixed table without requiring frequently repeated projective randomisation.

An additional advantage of the new method is that it is easily parallelizable on two-processor systems. One essential change in strategy compared with earlier methods for side-channel attack resistant point multiplication is the use of a right-to-left method (the scalar is processed starting at the least significant digit, cf. [15]) whereas the conventional methods work in a left-to-right fashion.

The method works in three stages, which are called initialisation stage, right-to-left stage, and result stage.

First there will be a high-level view of these stages before they are discussed in detail.

The method for computing eP is parameterized by an integer w≧2 and a digit set B consisting of 2^w integers of small absolute value such that every positive scalar e can be represented in the form

$e=\sum _{0\le i\le l}^{}{b}_i{2}^{\mathrm{wi}}$ using digits b₁ εB; for exampleB={0,1, . . . , 2^w−1}orB={−w^w-1, . . .,2^w-1−1}

A representation of e using the latter digit set can be easily determined on the fly when scanning the binary digits of e in right-to-left direction.

If e is at most n bits long (i.e. 0<e<2ⁿ), l=└n/w┘ is sufficient.

Let B′ denote the set {|b| |bεB} of absolute values of digits, which has at least 2^{w−1}+1 and at most 2^w elements. The point multiplication method uses #(B′)+1 variables for storing points on the elliptic curve in projective representation: Namely, one variable A_b for each bεB′, and one additional variable Q.

Let A_b^init denote the value of A_b at the end of the initialisation stage, and let A_b^sum denote the value of A_b at the end of the right-to-left stage. The initialisation stage sets up the variables A_b(bεB′) in a randomized way such that A_b^init ≠0 for each b, but

$\sum _{b\in B^{\prime }}^{}b{A}_b^{\mathrm{init}}=0$ (O denotes the point at infinity, the neutral element of the elliptic curve group.)

Then the right-to-left stage performs computations depending on P and the digits b₁, yielding new values A_b^sum of the variables A_b satisfying

$A_b^{\mathrm{sum}}=A_b^{\mathrm{init}}+\sum _{\underset{b_i=b}{0\le i\le l}}^{}{2}^{\mathrm{wi}}P-\sum _{\underset{b_i=-b}{0\le i\le l}}^{}{2}^{\mathrm{wi}}P$ for each bεB′. Finally, the result stage computes

$\sum _{b\in B^{\prime }-\left\{0\right\}}^{}b{A}_b^{\mathrm{sum}},$ which yields the final result eP because

$\begin{array}{c}\sum _{b\in B^{\prime }-\left\{0\right\}}^{}b{A}_b^{\mathrm{sum}}=\underset{\underset{0}{︸}}{\sum _{b\in B^{\prime }-\left\{0\right\}}^{}b{A}_b^{\mathrm{init}}}+\\ \sum _{b\in B^{\prime }-\left\{0\right\}}^{}b\left(\sum _{\underset{b_i=b}{0\le i\le l}}^{}{2}^{\mathrm{wi}}P-\sum _{\underset{b_i=-b}{0\le i\le l}}^{}{2}^{\mathrm{wi}}P\right)\\ =\sum _{0\le i\le l}^{}{b}_i{2}^{\mathrm{wi}}P=\mathrm{eP}.\end{array}$

The point multiplication method is a signed-digit variant of Yao's right-to-left method [15] (see also [16, exercise 4.6.3-9]) and [17, exercise 4.6.3-9]) and [18]) with two essential modifications for achieving resistance against side-channel attacks: The randomized initialisation stage is different; and in the right-to-left stage, the digit 0 is treated like any other digit.

DETAILED DESCRIPTION OF THE INVENTION

In the following the three stages are discussed in detail describing possible implementations.

The initialisation stage can be implemented as follows:

• • 1. For each bεB′-{1}, generate a random point on the elliptic curve and store it in variable A_b.
  • 2. Compute the point −Σ_{bεB′−{0,1}b Ab} and store it in variable A_i.
  • 3. For each bεB′, perform a projective randomisation of variable A_b.

The resulting values of the variables A_b are denoted by A_b^unit.

If the elliptic curve is fixed, precomputation can be used to speed up the initialisation stage:

The steps 1 and 2 should be run just once, e.g. during personalization of a smart card, and the resulting intermediate values A_b stored for future use.

These values are denoted by A_b^fix. Then only step 3 (projective randomisation of the values A_b^fix to obtain new representations A_b^init) has to be performed anew each time the initialisation stage is called for. The points A_b^fix must not be revealed; they should be protected like secret keys.

Generating a random point on an elliptic curve is straightforward. For each element X of the underlying field, there are zero, one or two values Y such that (X,Y) is the affine representation of a point on the elliptic curve.

Given a random candidate value X, it is possible to compute an appropriate Y if one exists; the probability for this is approximately ½ by Hasse's theorem.

If there is no appropriate Y, one can simply start again with a new X.

Computing an appropriate Y given X involves solving a quadratic equation, which usually (depending on the underlying field) is computationally expensive.

This makes it worthwhile to use precomputation as explained above.

It is also possible to reuse the values that have remained in the variables A_b,b≠1, after a previous computation, and start at step 2 of the initialisation stage.

To determine −Σ_{bεB′−{0,1}}b A_b in step 2, it is not necessary to compute all the individual products bA_b.

The following Algorithm can be used instead to set up A₁ appropriately if B′={0,1, . . . , β}, β≧2.

(Note that both loops will be skipped in the case β=2.)

Algorithm 1 Compute A1   −Σbε{2, ,β}bAb in the initialisation stage
for i = β − 1 down to 2 do
Ai   A1 + Ai+l
A1   2A2
for i = 2 to β − 1 do
Ai   Ai − Ai+l
A1   A1 + Ai+1
A1   −A1

This algorithm takes one point doubling and 3β-6 point additions.

When it has finished, the variables A_b for 1<b<β will contain modified values, but these are representations of the points originally stored in the respective variables.

If sufficient memory is available, a faster algorithm can be used to compute A₁ without intermediate modification of the variables A_b for b>1 (use additional variables Q_b instead; a possible additional improvement can be achieved if point doublings are faster than point additions).

The projective randomisation of the variables A_b(bεB′) in step 3 has the purpose to prevent adversaries from correlating observations from the computation of A₁ in the initialisation stage with observations from the following right-to-left stage. If algorithm 1 has been used to compute A₁ and the points are not reused for multiple invocations of the initialisation stage, then no explicit projective randomisation of the variables A_b for 1<b<β is necessary; and if β>2, no explicit projective randomisation of A₁ is necessary:

The variables have automatically been converted into new representations by the point additions used to determine their final values.

The following implements the right-to-left stage using a uniform pattern of point doublings and point additions.

Initially, for each b, variable A_b contains the value A_b^init; the final value is denoted by A_b^sum.

Algorithm 2 Right-to-left stage
Q   P
for i = 0 to l do
if bi ≧ 0 then
Abi   Abi + Q
else
A|bi|   A|bi| − Q
Q   2w Q

Due to special cases that must be handled in the point addition algorithm ([19]), uniformity of this algorithm is violated if A_|bi| is a projective representation of ±Q; the randomisation in the initialisation stage ensures that the probability of this is negligible.

(This is why in the section, where the initialisation stage is described, it is required that precomputed values A_b^fix be kept secret.)

If B contains no negative digits, the corresponding branch in the algorithm can be omitted.

The obvious way to implement Q←2^wQ in this algorithm is w-fold iteration of the statement Q←2 Q, but depending on the elliptic curve, more efficient specific algorithms for w-fold point doubling may be available (see [20]).

In the final iteration of the loop, the assignment to Q may be skipped (the value Q is not used after the right-to-left stage has finished).

With this modification, the algorithm uses lw point doublings and l+1 point additions. Observe that on two-processor systems the point addition and the w-fold point doubling in the body of the loop may be performed in parallel: Neither operations depends on the other's result.

Similarly to the computation of A₁ in the initialisation stage, the result stage computation

$\sum _{b\in B^{\prime }-\left\{0\right\}}^{}b{A}_b^{\mathrm{sum}}$ can be performed without computing all the individual products b _Ab^sum. In the result stage, it is not necessary to preserve the original values of the variables A_b, so the following algorithm (from [16, answer to exercise 4.6.3-9]) can be used if B′={0,1, . . . , β} when initially each variable A_b contains the value A^sum_b.

Algorithm 3 Compute Σbε{1, ,β}b Absum when initially Ab = Absum
for i = β − 1 down to 1 do
Ai   Ai + Ai+1
for i = 2 to β do
Al   Al + Ai
return Al

This algorithm uses 2β-2 point additions. Elliptic curve point arithmetic usually has the property that point doublings are faster than point additions. Then the variant described in the following algorithm is advantageous.

Algorithm 4 Compute Σbε{1,...,β}b Absum when initially Ab = Absum (variant)
for i = β down to 1 do
if 2i ≦ β then
Ai   Ai + A2i
if i is even then
if i < β then
Ai   Ai + Ai+1
Ai   2Ai
else
if i > 1 then
Ai   Al + Ai
return Al

This algorithm uses └β/2┘ point doublings and 2β−2−└β/2┘point additions.

Claims

1. A cryptographic method for performing an elliptic curve point multiplication eP where e is an integer and P is a point on an elliptic curve for use in cryptographic processing device, comprising the following steps: representing the multiplier e in the form

2. A cryptographic method for performing an elliptic curve point multiplication eP where e is an integer and P is a point on an elliptic curve for use in cryptographic processing device, comprising the following steps: representing the multiplier e in the form

3. A cryptographic method for performing an elliptic curve point multiplication eP where e is an integer and P is a point on an elliptic curve for use in cryptographic processing device, comprising the following steps: representing the multiplier e in the form

4. A cryptographic method for performing an elliptic curve point multiplication eP where e is an integer and P is a point on an elliptic curve for use in cryptographic processing device, comprising the following steps: representing the multiplier e in the form

5. The method according to claim 1, wherein in the modifying step the values 2wi P are computed in succession for i=0, . . . ,l and for each i the respective value is added to variable Abi.

6. The method according to claim 3, wherein in the modifying step the values 2wi P are computed in succession for i=0, . . . ,l and for each i the respective value is added to variable Abi if bi≧0 and subtracted from variable A−bi if bi<0.

7. The method according to claim 1, wherein 1εB, and wherein in the assigning randomly selected point representations step random points are assigned to the variable Ab with bεB−{1} and the negative of the value of the sum

8. The method according to claim 7, wherein the points in Ab are described in projective coordinates and wherein at the end of the assigning randomly selected point representations step a projective randomization takes place of the variables Ab with bεB−{1}.

9. A device comprising means, configured to provide the method according to the previous independent method claim 1.

10. A device comprising means, configured to provide the method according to the previous independent method claim 2.

11. A device comprising means, configured to provide the method according to the previous independent method claim 3.

12. A device comprising means, configured to provide the method according to the previous independent method claim 4.

13. A computer readable medium comprising computer executable instructions that when executed provide the method according to the previous independent method claim 1 while being executed on a computer.

14. A computer readable medium comprising computer executable instructions that when executed provide the method according to the previous independent method claim 2 while being executed on a computer.

15. A computer readable medium comprising computer executable instructions that when executed provide the method according to the previous independent method claim 3 while being executed on a computer.

16. A computer readable medium comprising computer executable instructions that when executed provide the method according to the previous independent method claim 4 while being executed on a computer.