Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

Elliptic Curves Cryptosystems (ECC) have been introduced by N. Koblitz [22] and V. Miller [29]. Their security relies on the hardness of the discrete logarithm problem. Elliptic curve based algorithms usually require keys far smaller than those involved in other public-key cryptosystems like RSA. This explains the current popularity of ECC and their involvement in a large variety of applications implemented over all kinds of devices: smart-cards, micro-controllers, and so on. Since such devices are widespread and in the hands of end-users, they are confronted to a wide range of threats. In particular, physical attacks need to be taken into account when assessing the overall security of the implementation. Thus, countermeasures are often conceived and implemented alongside the algorithms.

Physical attacks are traditionally divided into the two following families: perturbation analysis and observation analysis. The first one aims to modify the cryptosystem processing with any physical mean such as laser beams, clock jitter or voltage perturbation (e.g. fault injection attacks [12, 13]). The attacker then learns information on the secret parameter by observing the response of the cryptosystem to this perturbation. Such attacks can be prevented by monitoring the device environment with captors and by verifying the result of the computation before output. The second family of attacks consists in measuring physical data during the algorithm execution and then in exploiting this information to recover the secret. Such leakage sources can be the power consumption or the electro-magnetic emanation. Among observation attacks, we can distinguish several categories. Simple Power Analysis [24] directly deduces the value of the secret from a single input processing (possibly averaged), while Advanced Power Analysis exploits observations for several algorithm inputs. The latter kind of attacks requires the choice of a leakage model to compare predictions based on key-hypotheses with the measured traces. The comparison is done with a statistical tool, also called distinguisher. For example, the well-known Differential Power Analysis (DPA) [25] uses the difference of means whereas the Correlation Power Analysis (CPA) [10] involves Pearson’s correlation coefficient Footnote 1.

Countermeasures against observation analysis fall into two categories: hiding techniques aim at reducing the Signal to Noise Ratio (SNR) by increasing the noise or by equalizing the current in the circuit [35], while masking techniques consists in randomizing the sensitive computations. In practice, masking is always applied (possibly combined with hiding) since it provides strong security guaranty. Recently, Lee et al. [26] proposed a new efficient countermeasure to overcome first-order CPA Footnote 2 attacks. It assumes that the field operations are performed in the Montgomery domain [30] and consists in randomizing the Montgomery representation of the internal results (thus defining a so-called Randomized Montgomery Domain). This countermeasure has been considered as a very valuable alternative to the previous techniques because it avoids the need for scalar blinding (see e.g. [5]) and is much more efficient from a hardware implementation point of view.

Our work. In this paper, we show that the countermeasure proposed in [26] is flawed and can be efficiently broken by first-order CPA, even in presence of a large amount of noise in the measurements. After a presentation of the techniques proposed by Lee et al.  in Sect. 2, the flaw is exhibited in Sect. 3.1. The attack is afterwards detailed in Sect. 3.2 and its efficiency is demonstrated in Sect. 4 thanks to simulations in various contexts. Finally, Sect. 5 analyses the experiments given in [26] and concludes with a short discussion about possible countermeasures.

2 On Randomized Implementations of Modular Operations

In this section, we recall some mathematical background on elliptic curve and the associated scalar multiplication. The reader may also refer to [3] for a more complete overview. Here, we will focus on the efficient implementation of the scalar multiplication in embedded devices. In particular we recall the use of the Montgomery domain for efficient modular operations [31].

2.1 Background on Elliptic Curves and Montgomery Multiplication

Elliptic Curves. In this paper, we focus on elliptic curves \(E\) defined over a prime field \(\mathbb {F}_{p}\) according to the following short Weierstrass’s Equation:

$$\begin{aligned} E:y^2 = x^3+ax +b, \end{aligned}$$
(1)

where \(a\) and \(b\) are elements in \(\mathbb {F}_{p}\) satisfying \(4a^3+27b^2\ne 0\). The set of rational points of \(E\) is denoted by \(E(\mathbb {F}_{p})\). It contains all the points whose coordinates \((x,y)\in \mathbb {F}_{p}^2\) satisfy (1). This set, augmented with a neutral element \(\mathcal {O}\) called point at infinity, has an Abelian group structure for the following addition law: let \(P=(x_1,y_1)\) and \(Q=(x_2,y_2)\) then the coordinates \((x_3,y_3)\) of \(P+Q\) satisfy;

$$\begin{aligned} x_3 = \lambda ^2 - x_1 - x_2 \text { and } y_3 = \lambda (x_1-x_3) -y_1, \end{aligned}$$
(2)

where \(\lambda \) equals \((y_2-y_1)/(x_2-x_1)\) if \(P\ne Q\) and \((3x_1^2+a)/(2y_1)\) otherwise. The scalar multiplication (ECSM for short) of a point \(P\in E(\mathbb {F}_{p})\) by a natural integer \(k\) is denoted by \(kP\). It is the core operation of many cryptographic protocols such as ECDSA [22] and defining an efficient scalar multiplication arithmetic is hence a central issue (the interested reader is referred to [16] for a good overview). The point addition formula (2) is also central in elliptic curve implementations and several papers have been published on this subject [1, 4, 14, 15, 18, 27, 28, 41]. They aim at proposing sequences of operations over \(\mathbb {F}_{p}\) which are optimal according to some relations between the cost of a field inversion, the cost of an addition/subtraction and that of a multiplication. To defeat SPA attacks [5] at the arithmetic level, some works also try to propose sequences that stay unchanged whether (2) defines an addition (\(P\ne Q\)) or a doubling (\(P=Q\)). In this paper, we make no particular assumption on the type of ECSM algorithm nor on the sequence of operations which is used to process a point addition or a doubling. We indeed present a side-channel attack that does not exploit some particularity of the operations sequence or modus operandi but exploits information leakage during the manipulation of an intermediate result when the masking proposed in [26] is used. The latter type of result is likely to appear whatever the implementation choice.

Montgomery Domain. In order to efficiently perform the field operations involved in the point addition/doubling, developers often use the well-known Montgomery arithmetic [31]. This technique allows to replace the divisions occurring during the modular reduction by very efficient binary shifts. The only costly operation is the transformation to and from the Montgomery domain, but it is only done twice: at the beginning and at the end of the whole point operation (e.g.  the scalar multiplication).

Let \(x\) represent an element in the prime field \(\mathbb {F}_{p} \simeq \mathbb {Z}_{p}\) and let \(R = 2^m\) be defined such that \(2^m<p<2^{m+1}\) (i.e.  \(m=\lfloor \log _2(p)\rfloor \)). The value \(\overline{x}= x\cdot R \mathrm{{~mod~}}{p}\) is called the Montgomery representation of \(x\) and \(R\) is called the Montgomery constant. There is an isomorphism between \((\mathbb {Z}_{p},+,\cdot )\) and the Montgomery domain \(\mathrm {MD}(p) = (\{\overline{x}\},+,\otimes )\), where \(\otimes \) represents the Montgomery multiplication \(\overline{x}\otimes \overline{y} = (\overline{x}\cdot \overline{y})\cdot R^{-1} \mathrm{{~mod~}}{p}\). The result of the multiplication corresponds to \(\overline{xy}\), i.e. the Montgomery representation of \(xy \mathrm{{~mod~}}{p}\).

2.2 Randomized Montgomery Domain

Instead of using always \(R = 2^m\), the authors of [26] propose to use a randomized Montgomery constant \(r = 2^{\lambda }\) where \(\lambda \) is the Hamming Weight of a random \(m\)-bit value. The randomized Montgomery representation (RMR for short) of \(x\in \mathbb {Z}_{p}\) is \(x\cdot 2^{\lambda } \mathrm{{~mod~}}p\). It is denoted \(\widetilde{x}_{\lambda }\) or just \(\widetilde{x}\) when there is no ambiguity on \(\lambda \). The randomized Montgomery domain is denoted by \(\mathtt{RMD }_{\lambda }(p)\). The multiplication in \(\mathtt{RMD }_{\lambda }(p)\) works as follows: \(\widetilde{x}\otimes _{\lambda } \widetilde{y} = (\widetilde{x} \cdot \widetilde{y})2^{-\lambda } \mathrm{{~mod~}}{p}\). The result of the multiplication corresponds to \(\widetilde{xy}\), i.e. the RMR of \(xy \mathrm{{~mod~}}{p}\). To ensure that the number of final subtractions stays upper-bounded by \(1\) (as in the classical Montgomery multiplication), the random power \(\lambda \) must be chosen in \([0..m]\).

In [26], the idea of RMR is applied to secure the implementation of an elliptic curve scalar multiplication against first-order side channel attack (e.g. CPA). The principle of these attacks is to observe the device behaviour during the processing of several scalar multiplications \(kP\) where the secret scalar \(k\in \mathbb {N}\) stays unchanged and the public point \(P\) varies. Such attacks can be applied, for example, against some implementations of semi-static Diffie-Hellman key exchange, as found in the IEEE P1363 standard [17]. The main steps of the secure algorithm proposed in [26] are recalled hereafter, where the point \(P\) is assumed to belong to the set of rational points \(E(\mathbb {F}_{p})\) defined as in (1).

  1. 1.

    Inputs: a (public) point \(P=(x_{1},y_{1})\in E(\mathbb {F}_{p})\), a (secret) scalar \(k\in \mathbb {N}\), a random number \(\alpha \in [0..2^m-1]\) with \(m=\lfloor \log _2(p) \rfloor \).

  2. 2.

    Conversion to RMD : for \(\lambda = \mathrm {HW}(\alpha )\), process

    $$ \widetilde{x_{1}}= x_1\cdot 2^{\lambda } \mathrm{{~mod~}}p $$

    and

    $$ \widetilde{y_{1}}= y_1\cdot 2^{\lambda } \mathrm{{~mod~}}p. $$

    Also convert the curve parameters \((a,b)\) into \(\mathtt{RMD }_{\lambda }(p)\). The point with coordinates \((\widetilde{x_{1}},\widetilde{y_{1}})\) is denoted by \(\widetilde{P}\).

  3. 3.

    Elliptic Curve Scalar Multiplication ( ECSM ): in \(\mathtt{RMD }_{\lambda }(p)\), process

    $$\begin{aligned} \widetilde{Q}=(\widetilde{x_{2}},\widetilde{y_{2}}) = k \widetilde{P}. \end{aligned}$$
    (3)
  4. 4.

    Conversion to Integer Domain: process

    $$ x_{2} = \widetilde{x_2} \otimes _{\lambda } 1 = \widetilde{x_{2}}\cdot 2^{-\lambda } \mathrm{{~mod~}}p $$

    and

    $$ y_{2} = \widetilde{y_2} \otimes _{\lambda } 1 = \widetilde{y_{2}}\cdot 2^{-\lambda } \mathrm{{~mod~}}p $$
  5. 5.

    Output: \(Q = (x_{2},y_{2}) = kP\)

The Elliptic Curve Scalar Multiplication (ECSM) may be done with any algorithm (e.g. [31]). The authors of [26] do not recommend any particular one, even if the resistance tests reported in their paper are applied against an ECSM based on Montgomery Ladder (see Algorithm 5 in Appendix A). Similarly, our attack described in the next section does not exploit any particular feature of the ECSM and thus applies independently from the choice of algorithm. To allow for comparisons with [26] we however chose to target an ECSM based on Montgomery Ladder in our attack simulations.

3 Our Attack

3.1 Core Idea

The goal of our attack is to recover the bits of \(k\) one after another during the processing of ECSM in the randomized Montgomery domain (i.e. the processing of (3)). In this section, we detail how the second MSB (respectively the LSB) of \(k\) is recoveredFootnote 3. Once this bit is obtained, the attack can be applied similarly on the other bits from left to right (respectively right to left).

We denote by \(R(\widetilde{u},\widetilde{v})\) the value of an intermediate point during the processing of ECSM in the randomized Montgomery domain. We assume that the coordinates of the point \(R\) depend on a small sub part of \(k\) (e.g. a bit). This part is denoted by \(s\). It can for instance correspond to the result of the second point operation in ECSM with Montgomery Ladder (see (8)) or to the result of the third point operation in ECSM with double-and-add always method (see (6)–(7)). Our side channel attack targets the manipulation of the first coordinate \(\widetilde{u}\) of this point by the device: by assumption, it satisfies \(\widetilde{u} = f(\widetilde{x},\widetilde{y},s)\) where we recall that \((\widetilde{x},\widetilde{y})\) denotes the RMR coordinates of the input point \(P\) and where \(f\) is a known function that depends on the curve parameters and the algorithm used to process ECSM (see Appendix A for examples). To simplify the presentation, we assume that \(s\) is reduced to a single bit of \(k\) (the MSB or the LSB) but our attack straightforwardly applies to higher values (the only restriction is that the upper bound must be small enough to allow for an exhaustive test of all the elements). Our attack is based on the following statement which essentially means that a RMR representation leaks information on the un-blinded coordinate and that this information can be exploited by a first-order side-channel attack:

Statement: the function \(f:x\mapsto \mathbb {E}\big [\mathrm {HW}(\widetilde{X}_{\lambda }) ~ | ~ X=x\big ]\) is not constant.

Remark 1

Taking into account the specificities of the randomized Montgomery representation recalled in Sect. 2.2, the mean is computed over the random variable \(\lambda \) defined such that \(\lambda = \mathrm {HW}(\alpha )\) with \(\alpha \) having a uniform distribution on \([0..2^m-1]\).

To argue on the statement above, we evaluated \(f\) on \(1000\) different values \(x\) for \(m=256\) and \(m=384\) (these parameters are recommended for instance by the American National Security Agency when ECSM is used to process an ECDSA). The results are plotted in Fig. 1(a) and (b). For comparison, we also plotted in Fig. 2(a) and (b) the value \(\mathrm {HW}(x)\) for \(1000\) values \(x\) in \([0..p)\), which corresponds to the case of non blinded values. Finally, in Fig. 3(a) and (b) we plotted \(\mathbb {E}\big [\mathrm {HW}(\lambda \cdot X \mathrm{{~mod~}}{p}) ~ | ~ X=x\big ]\) where \(\lambda \) is a random uniformly distributed \(16\)-bit value. The latter corresponds to the case where ECSM is protected by blinding the projective coordinates of the point \(P\).

Fig. 1.
figure 1

Values of \(f:x\mapsto \mathbb {E}\big [\mathrm {HW}(\widetilde{X}) ~ | ~ X=x\big ]\) for 1000 random values \(x\)

Full size image
Fig. 2.
figure 2

Values of \(f:x\mapsto \mathrm {HW}(x)\) for 1000 random values \(x\)

Full size image
Fig. 3.
figure 3

Values of \(f:x\mapsto \mathbb {E}\big [\mathrm {HW}(\lambda \cdot X \mathrm{{~mod~}}{p}) ~ | ~ X=x\big ]\) for 1000 random values \(x\)

Full size image

As expected, we can see in Fig. 2(a) and (b) that the Hamming weight of \(x\) varies with \(x\) when no blinding is involved; this implies that a first-order CPA is possible if \(x\) is sensitive. On the contrary, Fig. 3(a) and (b) show that coordinate blinding cancels any leakage on \(x\) since the average Hamming weight of the blinded coordinate is almost constant (even in our case where we limited \(\lambda \) to \(16\)-bits words which can be considered as too limited in practice). Between the two previous extreme cases, Fig. 1(a) and (b) show that \(f(x)\) varies with \(x\) (not with a high variance as in Fig. 2(a) and (b) but visibly much more than in Fig. 3(a) and (b)): this implies that the countermeasure in [26] can be attacked with a first-order CPA involving \(f\) to process the predictions on \(\widetilde{x}\). In the following section, we detail the latter attack.

3.2 Attack Description

As explained in the previous section, our attack recovers the secret piece by piece. Here, we detail how it allows for the recovery of the second MSB or the LSB \(s\) of the secret scalar \(k\) (the principle can then be repeated to recover the other bits one after another). The intermediate result/point \(R(\widetilde{u},\widetilde{v})\) exploited by the attack differs with the algorithm used to process the scalar multiplication. We give some examples below but our first-order side-channel attack also applies in other contexts (e.g. against atomic implementations or implementations applying the window principle [21]) as long as the point is blinded with the RMR representation proposed in [26] and the scalar is not blinded.

Several scalar multiplication algorithms are recalled in Appendix A. We recall that, for left-to-right versions, we make the (classical) assumption that the MSB equals \(1\). Our attack targets the manipulation of the first coordinate \(\widetilde{u}\) of the intermediate result/point \(R(\widetilde{u},\widetilde{v})\) corresponding to:

  • [left-to-right double-and-add ECSM ] the second point operation

    $$\begin{aligned} R=3P+(1-s)P. \end{aligned}$$
    (4)

  • [right-to-left double-and-add ECSM ] the first point operation

    $$\begin{aligned} R=P+(1-s)P. \end{aligned}$$
    (5)

  • [left-to-right double-and-add-always ECSM ] the third point operation

    $$\begin{aligned} R=2(P+sP). \end{aligned}$$
    (6)

  • [right-to-left double-and-add-always ECSM ] the third point operation

    $$\begin{aligned} R=2P+sP. \end{aligned}$$
    (7)

  • [Montgomery Ladder ECSM ] the second point operation

    $$\begin{aligned} R=2(P+sP). \end{aligned}$$
    (8)

  • [Joye ECSM ] the second point operation

    $$\begin{aligned} R=P+(1-s)P. \end{aligned}$$
    (9)

Following the classical outlines of a first-order side-channel attack, our attack starts by the observation of the device behaviour for several executions of the algorithm ECSM parametrized by different public inputs \(P\) but a same secret scalar \(k\). Note that only the part of the observation corresponding to the manipulation of the coordinate \(\widetilde{u}\) of \(R\) is used in the attack described hereafter. Since the \(i\) \(^{\text {th}~}\)observation (e.g. the power consumption or the electromagnetic emanation) is algorithmically related to the \(i\) \(^{\text {th}~}\)input point \(P_i\) and the secret bit \(s\), it is denoted by \(\mathcal {L}(s,P_i)\) in the following. At the end of this measurement step of the attack, the adversary is assumed to be provided with a sample of pairs \(\left( P_i,\mathcal {L}(s,P_i)\right) _i\). The size of this sample is denoted by \(N\).

To underline the functional dependency between \(\widetilde{u}\) and \((s,P_i)\), we use the notation \(\widetilde{u}(s,P_i)\) in the following. For testing an hypothesis \(\hat{s}\) on \(s\), the attack continues with the computation of the values \(u(\hat{s},P_i)\) for \(i\in [1..N]\). These values correspond to the unmasked version of the \(u\)-coordinates under the hypothesis \(\hat{s}=s\). Then applying the strategy already used in [6, 36] and argued in [38], we choose a device model \(\mathsf {m}\) (usually the Hamming weight) and we compute the sample of predictions \((h_i)_{i\leqslant N}\) defined such that:

$$ h_i(\hat{s}) = \mathbb {E}_{\lambda } \big [ \mathsf {m}(u(\hat{s},P_i)\cdot 2^{\lambda } \mathrm{{~mod~}}p)\big ]= \sum _{i=1}^m \mathsf {m}(u(\hat{s},P_i)\cdot 2^{i} \mathrm{{~mod~}}p) \times \mathrm {p} [\lambda = i]. $$

By construction \(\lambda \) is defined as the Hamming weight of an \(m\)-bit random element. Its distribution is therefore binomial with parameters \(m\) and \(1/2\) and the equation above can be developed as follows:

$$\begin{aligned} h_i(\hat{s}) = \frac{1}{2^m} \sum _{i=1}^m \mathsf {m}(u(\hat{s},P_i)\cdot 2^{i} \mathrm{{~mod~}}p) \times \left( {\begin{array}{c}m\\ i\end{array}}\right) . \end{aligned}$$
(10)

Eventually, the absolute value of the correlation \(\rho _{\hat{s}}\) between the samples \((h_i(\hat{s}))_i\) and \((\mathcal {L}(s,P_i))_i\) is computed for \(\hat{s}\in \{0,1\}\) and the attack returns the hypothesis with the greatest correlation.

4 Simulations

Setting. In this section, we assume that the ECSM is implemented on a \(32\)-bit architecture and according to the Montgomery ladder (Algorithm 5 in Appendix A). In such environment which corresponds to a classical context, data will only be manipulated through \(32\)-bit registers. Without loss of generality, we hence assume that the leakage exploited in our attacks is the Hamming weight of the \(32\) least significant bits of \(\widetilde{u}\) (namely \(\mathrm {HW}(\widetilde{u} \mathrm{{~mod~}}2^{32})\)) instead of the Hamming weight of the whole \(256\) or \(384\) bit valueFootnote 4. The described attack aims at recovering the most significant bit \(s\) of the secret \(k\) and, according to (8), the observations are hence assumed to be related to the manipulation of the coordinate \(\widetilde{u}\) of the point \(R(\widetilde{u},\widetilde{v})=2(P+sP)\). The latter observations are simulated in the classical Hamming weight model with Gaussian Noise. Namely, the leakage observation \(\mathcal {L}(s,P_i)\) corresponding to the \(i\) \(^{\text {th}~}\)scalar multiplication is simulated such that:

$$\begin{aligned} \mathcal {L}(s,P_i) = \mathrm {HW}(\widetilde{u} \mathrm{{~mod~}}2^{32}) + \mathcal {N}, \end{aligned}$$
(11)

where \(\widetilde{u}=u\cdot 2^{\lambda _i}\mathrm{{~mod~}}p\) is the first coordinate of \(2(P_i+sP_i)\) represented in the randomized Montgomery domain \(\mathtt{RMD }_{\lambda _i}(p)\) (with \(\lambda _i\) generated at random in its definition set) and where \(\mathcal {N}\) is a Gaussian random variable with mean \(0\) and standard deviation \(\sigma \).

Since the leakage is assumed to satisfy (11), the model function \(\mathsf {m}\) in Eq. (10) has been simply chosen to be the Hamming weight of the \(32\) least significant bits of the input. The predictions \(h_i(\hat{s})\) associated to the binary hypothesis \(\hat{s}\) on \(s\) hence satisfy:

$$\begin{aligned} h_i(\hat{s}) = \frac{1}{2^m} \sum _{i=1}^m \mathrm {HW}((u\cdot 2^{i} \mathrm{{~mod~}}p) \mathrm{{~mod~}}2^{32}) \times \left( {\begin{array}{c}m\\ i\end{array}}\right) , \end{aligned}$$
(12)

where \(u\) is the first coordinate of \(2(P_i+\hat{s}P_i)\) in the standard integer domain and where \(m\) equals \(256\) or \(384\).

To sum-up, the simulations reported in this section are split in two parts. The first part aims to show that the CPA (i.e. the correlation coefficient) succeeds in distinguishing the correct hypothesis from the wrong one. The second part serves to estimate the success rate in recovering the first secret bit in a Montgomery ladder implementation of ECSM. The latter success rate is estimated for different noise levels \(\sigma \) and different number of observations.

Fig. 4.
figure 4

Correlation for 256-bit curve attacking with known messages

Full size image

Pearson Correlation Coefficient. First, we are comparing the correlation values for the correct and the wrong predictions. Following the attack description in Sect. 3.2, we compute the set of hypotheses \((h_i(\hat{s}=1))_{i\leqslant N}\) according to (12), and the two sets of leakages \((\mathcal {L}(1,P_i))_{i\leqslant N}\) and \((\mathcal {L}(0,P_i))_{i\leqslant N}\), according to (11), where we recall that \(N\) represents the total number of randomly chosen points \(P_i\). The first set corresponds to leakages that match the chosen hypotheses and the second one to leakages that do not match the hypotheses. These experiments have been repeated for different values of \(N \in [500..10000]\) and different noise standard deviation \(\sigma \in [0..30]\). Then, we compute the two correlation coefficients \(\rho _{\small {\hbox {correct}}} = \rho ((h_i(\hat{s}=1))_i , (\mathcal {L}(1,P_i))_i)\) and \(\rho _{\small {\hbox {wrong}}}=\rho ((h_i(\hat{s}=1))_i , (\mathcal {L}(0,P_i))_i)\). The obtained correlation values are then averaged over the execution of \(100\) such attacks and the standard deviation of the correlation values is also computed to measure how much the mean is informative.

Results of this first set of experiments are presented in Fig. 4. It represents the average and the standard deviation cone for the correlation coefficients obtained with the correct and wrong predictions. It may be seen that the correlation coefficient allows us to distinguish a correct key bit hypothesis from a wrong one with high probability as long as the number of observations satisfies:

$$ \left\{ \begin{array}{lcl} N \ge 500 &{} \hbox { if } &{} \sigma =0\\ N \ge 1500 &{} \hbox { if } &{} \sigma = 10\\ N \ge 3000 &{}\hbox { if } &{} \sigma = 20\\ N \ge 6000 &{} \hbox { if } &{} \sigma = 30\\ \end{array} \right. . $$
Fig. 5.
figure 5

Success rate for several noise values

Full size image

Success Rate. We then proceed to test the attack efficiency. For such a purpose, we randomly choose the secret \(s\) we are trying to recover. We then simulate the leakage observations according to (11) for \(N\) pairs \((P_i,\lambda _i)\) of values generated at random in their respective definition set. This step provides us with a set \((\mathcal {L}(s,P_i))_{i\leqslant N}\) playing the role of the registered traces in a real attack scenario. We then compute the two sets of predictions \((h_i(\hat{s}=0))_{i\leqslant N}\) and \((h_i(\hat{s}=1))_{i\leqslant N}\), and we process the absolute value of the two corresponding correlation coefficients \(\rho _0=\rho ((\mathcal {L}(s,P_i)_i, (h_i(\hat{s}=0))_i)\) and \(\rho _1=\rho ((\mathcal {L}(s,P_i)_i, (h_i(\hat{s}=1))_i)\). The prediction \(\hat{s}\) such that \(|\rho _{\hat{s}}|\) has the highest value is set as the most likely one. If \(\hat{s}=s\) the attack succeeds, else it fails. In our experiments we repeated each attack \(1000\) times with different random values to build a success rate.

Figure 5 represents the success rate in percentage when trying to guess the bit \(s\). In abscissa is the number of different inputs used for the computation of the correlation coefficient. Several noise standard deviation values have been used. It may be checked that even for \(\sigma =40\) (i.e. SNR \(=0,005\)) the attack succeeds with probability greater than \(80\,\%\) if the number \(N\) of observations used by the attacker is greater than \(5500\) for curves of size \(256\) bits, respectively greater than \(7000\) for \(384\)-bit curves.

5 Analysis and Conclusion

In this paper, we have argued that the countermeasure proposed in [26] is flawed and does not defeat first-order CPA. Actually, through attack simulations conducted under reasonable and classical assumptions on the execution environment we have shown that a CPA is likely to be very efficient even when the noise is huge. The problem identified in the countermeasure under study is that the distribution of the masking values \(\lambda \) is binomial (and not uniform). This choice, which has been done for efficiency reason, is in fact dramatic from a security point of view. It must be mentioned that the authors of [26] tested a school-book CPA against their implementation and used the failure of this attack to argue on the resistance of their countermeasure. As we have shown here, the failure of the CPA performed in [26] is not a consequence of the countermeasure quality but of a wrong attack parametrization. By adapting to a first-order context the argumentation given in [38] for second-order attacks, we were indeed able to use a much better parametrization for the CPA. Our work demonstrates once again that countermeasures must not be only validated by performing some ad-hoc attacks (even classical) but must be formally analysed, for instance by following the approaches proposed in [7, 9, 37].

Concerning the proposed countermeasure, a possible patching could consist in generating \(\lambda \) uniformly in \([0..m]\) but a careful security analysis is needed to assess on the pertinence of this patch. The proposal of Dupaquis and Venelli in [8] may offer an interesting alternative in the case where efficiency is required. Otherwise, well-studied approaches such as classical exponent and message blinding seem to offer better security garanties.