Cryptographic Divergences: New Techniquesand New ApplicationsMarc Abboud1⋆ and Thomas Prest21 École Normale Supérieuremarc.abboud@ens.fr2 PQShieldthomas.prest@pqshield.comAbstract In the recent years, some security proofs in cryptography haveknown significant improvements by replacing the statistical distance withalternative divergences. We continue this line of research, both at a the-oretical and practical level. On the theory side, we propose a new cryp-tographic divergence with quirky properties. On the practical side, wepropose new applications of alternative divergences: circuit-private FHEand prime number generators. More precisely, we provide the first for-mal security proof of the prime number generator PRIMEINC [9], andimprove by an order of magnitude the efficiency of a prime number gener-ator by Fouque and Tibouchi [17,18] and the washing machine techniqueby Ducas and Stehlé [16] for circuit-private FHE.1 IntroductionCryptographic divergences play an essential role in cryptography. Most of thetime, they provide rigorous theoretical tools to prove that the concrete instan-tiation of a cryptosystem is as secure as an idealized description. Typically, theidealized scheme will rely on an ideal distribution Q, whereas its instantiationwill rely on a distribution P. If Div(P;Q) is small for some divergence Div, thenone can predict the security of the concrete cryptosystem based on the secu-rity of the ideal one. Similarly, cryptographic divergences can help to connect acryptosystem to a hard problem.The statistical distance is by far the most prevalent cryptographic divergence.It is simple and versatile, making it the swiss army knife of cryptography. How-ever, these last years have seen a number of works using alternative divergences.A compelling example is the Rényi divergence, which use has been spearheadedby lattice-based cryptography to improve security reductions [28,31,3,7,41,4].When the number of queries is limited and in the presence of a search prob-lem, it can provide significant gains. Another example is the Kullback-Leiblerdivergence, which has been used at a more foundational level: recent works have⋆ Most of this work was done while Marc Abboud was an intern at PQShield.2leveraged it to (re-)define fundamental notions such as the advantage of an adver-sary [34] or computational entropies [1], simplifying proofs or resolving paradoxesin the process1 .1.1 Our ContributionsIn this work, we continue the exploration of alternative divergences to improvesecurity proofs. This is done at two levels: by providing new theoretical tools,and by finding new applications to specialized divergences.A New Cryptographic Divergence. Our first contribution is to propose aparametric divergence called REα divergence (REα stands for relative error oforder α). We believe it is interesting for theoretic and practical reasons. On thetheory side, it connects several divergences: it is a continuous trade-off betweenthe statistical distance and the relative error, and enjoys desirable propertiesof both divergences. It is also (non-tightly) equivalent to the Rényi divergence.On the practical side, it satisfies an unusual amplification property with crypto-graphic applications. Relations between divergences is given in Figure 1.∆ML∆SD REα ∆REDKL Rα R∞≃∆SD ≤√12DKL (3)(4) R∞ ≤ 1 +∆RE2∆SD = RE1 ≤ REα ≤eDKL = R1 ≤ Rα ≤Lemma 3Figure 1: Relations between cryptographic divergences: max-log distance ∆ML,statistical distance ∆SD, REα divergence, relative error ∆RE, Kullback-Leiblerdivergence DKL, Rényi divergence Rα.New Applications: Proof Outline. As new applications of our new tech-niques and of existing ones, we provide improved analyses for circuit-privateFHE and prime number generators. Our proofs follow this blueprint – alreadyimplicit in [41]:(a) Bound the relative error ∆RE(R∥I) between a real distribution R and idealdistribution I;(b) Deduce the Rényi divergence Rα(R∥I) between R and I;1 [34] use the mutual information for its redefinition. The mutual information betweentwo random variables X,Y is the KL divergence between the their joint and productdistributions: MI(X;Y ) = DKL((X,Y );X × Y ).3(c) Conclude that an adversary making Q queries to I and trying to solve asearch problem does not increase his advantage by more than O(1) whenreplacing I by R.Our proofs follow either the logical structure (a)⇒ (b)⇒ (c), or (b)⇒ (c). Thejustification for (a) ⇒ (b) is given by Lemma 3, and the one for (b) ⇒ (c) byLemma 4. We emphasize that using the Rényi divergence instead of the statisticaldistance does not mean we prove something weaker or different; both divergencesare merely tools in proofs strategies, and we compare our improved analyses withexisting ones in the exact same setting (search problem, Q queries).Applications We provide two applications of our techniques: circuit privacyfor FHE and prime number generators. As is now customary when using theRényi divergence (and this is also true for the REα divergence), two conditionsare required to fully exploit the Rényi divergence:– The number of queries Q should be much lower than 2λ, where λ denotes thesecurity level. In practice 128 ≤ λ ≤ 256. On the other hand NIST’s call forpost-quantum cryptography standards suppose Q ≤ 264, and we may assumeeven lower bounds for computation- and bandwidth-heavy primitives suchas FHE. Finally, when generating a single public key, the number of queriesto the key generation algorithm is as small as 1 (in the single-target setting).– The underlying problem should be a search problem. One of our applicationstargets RSA-based signatures, where this is obviously the case. We also be-lieve that most practical usecases of circuit-private FHEs can be describedin a satisfying way with a search problem.In particular, we do not claim improved analyses for unlimited queries or decisionproblems. Unfortunately, in our practical usecases, the REα divergence (withα < ∞) does not give better results than using the relative error but we haveidentified theoretical situations where ∆RE can’t be used directly (see Section5.1).Application to Prime Number Generators. The ability to securely gener-ate prime numbers is essential for RSA-based cryptosystems. However, if primenumbers are sampled from a weak distribution, it can lead to a variety of attacks.The most common ones are GCD attacks, where an attacker collects RSA publickeys Ni = pi · qi with low collision entropy and extracts private keys (pi, qi) bycomputing GCDs. This multi-target attack has plagued the last decade, withseveral papers [25,29,5] compromising a total of more than 1 million keypairs.Primes sampled from highly structured distributions may also be vulnerable, asdemonstrated by Coppersmith’s attack [11,10] and its follow-up ROCA [39].To mitigate attacks, several algorithms have been proposed to securely gen-erate prime numbers. One obviously desirable property is to sample from adistribution with high collision entropy, lest the generated primes be vulnerableto GCD attacks. However, we note that having a high collision entropy does4not preclude Coppersmith’s attack, so it is not a necessary and sufficient con-dition for security. To offer stronger security guarantees, some prime numbersgenerators sample statistically close to the uniform distribution over primes in[2d; 2d+1].2 Some schemes [26,50] based on the strong RSA assumption explicitlyrequire this. In this work, we focus on two prime number generators and providesubstantially improved security proofs for them.The PRIMEINC generator. A prominent generator is PRIMEINC, proposedby Brandt and Damgård [9]. Due to its simplicity and entropy efficiency, it iscommonly used; see the PyCrypto3 and OpenSSL4 libraries. Despite its longevityand prevalence, PRIMEINC’s concrete security had remained an open question.Circumstantial arguments were presented, leaning either towards weak securityguarantees [9,36] or suggesting potential weaknesses [17,18], but no definite an-swer had been presented.We clarify the situation by providing formal arguments which guaranteePRIMEINC’s security in clearly defined scenarii or against common attacks.Our work for PRIMEINC uses only the Rényi divergence. More precisely, weshow that:• In the single-target setting, any scheme that is secure with the uniformdistribution U (over primes in [2d; 2d+1]) remains secure when replacing Uby the output distribution P of PRIMEINC, as long as there are O(1) callsto P; this covers for example RSA key generation. This argument is tight(only O(1) bits of security are lost) and fully generic.• In the multi-target setting, PRIMEINC has enough collision entropy to besecure against GCD attacks.The Fouque-Tibouchi generator. Fouque and Tibouchi [17,18] proposed primenumber generators with an appealing feature; the statistical distance betweentheir output distribution P and the uniform distribution U (over primes in an[2;x]) is upper bounded by log(x) · x−ϵ/4, where ϵ is an input parameter, and itcan therefore be proven arbitrarily close to 0 by increasing ϵ. These generatorsare provably secure. However, the entropy consumption is linear in ϵ: thus thereis a trade-off between statistical closeness to U and the entropy consumption. Astandard statistical distance argument would mandate ϵ ≥ 4λ log log xlog x , where λ isthe security level.We provide a Rényi divergence-based security argument that only mandatesϵ ≥ 2 log(λQ) log log xlog x , where Q is the number of queries to the generator. Theentropy gain is significant when Q is much smaller than 2λ, which is always thecase in real applications. For practical usecases, we gain an order of magnitude inentropy consumption. Our proof uses the REα divergence for the computations.2 Typically, additional requirements are mandated, such as (p+1) and (p− 1) havinga large prime factor; but these can be added on top of the sampling procedure.3 https://github.com/dlitz/pycrypto/blob/master/lib/Crypto/Util/number.py4 https://github.com/openssl/openssl/blob/master/crypto/bn/bn_prime.c5Application to Circuite-Private FHE. (Fully) homomorphic encryption –or (F)HE – allows to securely evaluate circuits on encrypted data. FollowingGentry’s breakthrough [21], it has known an exponential growth in the lastdecade, and is now being advertised as a product by companies (Duality, Inpher,Zama) and standardized.Circuit privacy is an increasingly relevant security notion for FHE. The set-ting is the following: a client C sends a (fully homomorphic) ciphertext c =Enc(m) to a server S, which then homomorphically computes c′ = Enc(f(m))for some function f , and sends it back to C. A standard security requirementis that S doesn’t learn anything about m or f(m). Conversely, circuit privacyrequires that C doesn’t learn anything about the circuit f in the process. Circuitprivacy is useful when f is a secret intellectual property of S; without circuitprivacy, a user might learn f and set up his own server. The lack of circuit pri-vacy can be a strong deterrent for a company wishing to provide its services onencrypted data.Today’s most efficient method to realize circuit privacy in a generic way isthe washing machine technique by Ducas and Stehlé [16].5 In a nutshell, it firstbootstraps the ciphertext, then injects entropy. One iteration of this bootstrap-then-inject-entropy process is called a cycle. [16] prove that their technique ensurecircuit privacy if Θ(λ) cycles are sequentially applied to the ciphertext, whereλ is the security level. However, despite recent improvements, bootstrappingremains an expensive operation. Thus circuit privacy can be a computationalbottleneck.We provide an improved analysis of Ducas and Stehlé’s washing machinetechnique. We reduce the number of cycles by a factor essentially 2λlogQ , where Qis the total number of (evaluation of f) queries made to the server. For realisticparameters, our new analysis improves the one of [16] by an order of magnitude.At a technical level, our proof leverages our proxy amplification property.1.2 Related WorksThese last years have seen a surge of papers using other divergences than thestatistical distance in the cryptographic literature. For example, the Hellingerdistance has been used to study key-alternating ciphers [46], the χ2 divergenceto study a few symmetric-key constructions [14], and the max-log distance [34]in the context of lattice-based cryptography.The Kullback-Leibler divergence has been used to improve parameters inlattice-based cryptography [40,15], to redefine the advantage [35], to unify com-putational entropy notions [1], and indirectly (via the mutual information) inside-channel analysis [23].The Rényi divergence has several applications in lattice-based cryptogra-phy [28,31,3,47,7,41,4]. Differential privacy [37,32] and leakage-resilient cryptog-raphy [42] have also benefitted from its use.5 The work of [8] requires no bootstrapping, but only applies to GSW-based schemesand is restricted to NC1.6Acknowledgements. The authors are indebted to Takahiro Matsuda andShuichi Katsumata for their insightful discussions and for pointing out a flaw inan earlier version of the paper. Thomas Prest is supported by the Innovate UKResearch Grant 104423 (PQ Cybersecurity).2 PreliminariesAsymptotic notations. For asymptotics, we use Laudau’s notation. For two realfunctions f, g : R+ → R, we note f = O(g) if there exists a constant C such that|f | ≤ C · |g|. Similarly, we note f = o(g) if f = ϵ · g for some function ϵ suchthat ϵ(x) −−−−→x→∞0. We note f = Θ(g) if f = O(g) and g = O(f). If f and g haveseveral variables, we note f = Ox(g) to specify that the assertion holds for thevariable x (and similarly for the other notations). We will also use the notationf ≪x g for f = Ox(g) when the articles we cite use it.Integers (modulo n). Z denotes the set of integers. Let n be an integer, Znwill denote the set of integers modn and Z×n will denote the group of invertibleelements of Zn. Euler’s totient function is φ(n) = |Z×n |.2.1 f-Divergencesf -divergences were first introduced by Csiszar [13], Morimoto [38] and Ali-Sirvey [2]. They provide a wide class of divergences between distributions, andencompass several divergences used in cryptography.Definition 1 (f-Divergences). Let f : R→ R be a convex function such thatf(1) = 0. Let P,Q be two distributions over a countable space X such thatSuppP ⊂ SuppQ. The f -divergence between P and Q is:Divf (P;Q) := EQ[f(PQ)]=∑x∈SuppQQ(x)f(P(x)Q(x))Special cases of f -divergences are the following:– Statistical distance: ∆SD(P;Q) = Divf (P;Q) for f : x 7→ 12 |x− 1|;– Kullback-Leibler divergence: DKL(P;Q) = Divf (P;Q) for f : x 7→x ln(x);– χ2 divergence: χ2(P;Q) = Divf (P;Q) for f : x 7→ (x− 1)2;We note that all f -divergences satisfy a few cryptographically useful prop-erties such as the data processing inequality, probability preservation propertiesand joint convexity (see resp. Lemma 1, Corollary 2 and Corollary 3 of [22]). Westate them here.Lemma 1 (Data-processing inequality, Lemma 1 [22]). Let P,Q be twodistributions over a space X and let T be a random function over X. Denote byPT ,QT the composition of respectively P and Q with T , thenDivf (PT ;QT ) ≤ Divf (P;Q).7Corollary 1 (Joint convexity of Divf , Corollary 2 [22]). All f -divergencesare jointly convex, i.e, for all distributions P1,P2,Q1,Q2 over a space X andfor all λ ∈ [0, 1], one hasDivf ((1−λ)P1+λP2; (1−λ)Q1+λQ2) ≤ (1−λ)Divf (P1,Q1)+λDivf (P2,Q2).Two divergences that are not f -divergences are also used in cryptography. IfSuppP = SuppQ = Ω, we define the following divergences:– Relative error: ∆RE(P;Q) = maxΩ∣∣PQ − 1∣∣;– Max-log distance: ∆ML(P;Q)maxΩ | lnP − lnQ|;Around 0, both divergences are equivalent [34]: ∆RE(P;Q) ∼ ∆ML(P;Q).2.2 Rényi Divergences and Rényi EntropiesIn this section, we recall the definitions of Rényi divergences and entropies.Definition 2 (Rényi divergence). Let P,Q be two discrete distributions overa space X such that SuppP ⊂ SuppQ. The Rényi divergence of order α is:Rα(P ;Q) :=(∑x∈SuppQP(x)αQ(x)α−1) 1α−1 if 1 < α <∞maxx∈SuppQP(x)Q(x) if α =∞eDKL(P;Q) if α = 1Note that Rα is not an f -divergence. However, Rα−1α − 1 is an f -divergencefor f : x 7→ xα − 1, which allows it to indirectly benefit from f -divergenceproperties. We recall some properties.Lemma 2 ([49,3]). For two distributions P,Q and two families of distributions(Pi)i, (Qi)i, the Rényi divergence verifies these properties:– Monotonicity. α ≥ 1 7→ Rα(P;Q) is a continuous non-decreasing function.– Data processing inequality. For any (randomized) function f , one hasRα(f(P); f(Q)) ≤ Rα(P;Q).– Multiplicativity. Rα(∏i Pi;∏iQi) =∏i Rα(Pi;Qi).– Probability preservation. For any event E ⊆ Supp(Q) and α ∈ (1,+∞),Q(E) ≥ P(E)αα−1 /Rα(P;Q), (1)Q(E) ≥ P(E)/R∞(P;Q). (2)The following lemma bounds the Rényi divergence from the relative error∆RE.Lemma 3 ([41]). Let P,Q be two distributions of same support such that∆RE(P;Q) ≤ δ. Then, for a ∈ (1,+∞):Ra(P;Q) ≤(1 +a(a− 1)δ22(1− δ)a+1) 1a−1∼δ→01 +aδ228We give a second lemma; it is already used implicitly in [41], but we makeit formal here as it is repeatedly used in our security arguments. In particular,given Lemma 4 initial conditions, no more than 1 + log2 3 ≤ 3 bits of securityare lost when switching from I to P.Lemma 4. Let A be an adversary performing Q queries to a distribution Iand solving a search problem SI with probability at most 2−λ, where λ > 1. Ifone replaces I with a distribution P such that Rλ(P; I) ≤ 1 + 1/Q, then theprobability that A solves SP is no more than 2−(λ−1) · e.Proof. Let E be the event of A solving S. Applying (1), then the multiplicativityof the Rényi divergence, and finally the inequality (1 + x/n)n ≤ ex yields:P(E)λ/(λ−1) ≤ I(E) ·Rλ(PQ; IQ) ≤ 2−λ ·Rλ(P; I)Q ≤ 2−λ · e.Therefore P(E) ≤ 2−(λ−1) · e(λ−1)/λ ≤ 2−(λ−1) · e.Rényi divergences have entropy measure counterparts, called Rényi entropiesor α-entropies. These have countless cryptographic applications [45].Definition 3 (Rényi entropy). Let α ∈ [1,+∞] and X be a discrete distribu-tion. The α-entropy (or Rényi entropy) of X is:Hα(X) =−∑xP[X = x] log2 P[X = x] if a = 111− αlog2(∑xP[X = x]α)if 1 < a <∞−maxxlog2 P[X = x] if a =∞H1 is also called Shannon’s entropy, H2 the collision entropy and H∞ the min-entropy. If we note U the uniform distribution over any superset of Supp(X),then all the Rényi entropies of U are equal: Hα(U) = log2 |Supp(X)| for anyα ∈ [1,+∞]. Note that Rényi divergences and entropies are closely related:log2 Rα(X;U) = Hα(U)−Hα(X).3 Theoretical ResultsIn this section, we present our theoretical results. We introduce a new crypto-graphic divergence; this parametric divergence seems deeply connected to severalexisting divergences, and possesses an unusual amplification property.93.1 A Parametric Divergence with Peculiar PropertiesWe now introduce a new divergence, which we note REα (for relative error) asit can be seen as a trade-off between the statistical distance ∆SD and the relativeerror ∆RE. This trade-off is parametrized by a scalar α ∈ [1,∞], and allows REαto be defined in situations where ∆RE is not, all the while sharing with ∆RE and∆SD a new desirable cryptographic properties.Definition 4 (REα divergence). Let P,Q be two distributions over a count-able space X such that SuppP ⊂ SuppQ. One defines the REα-divergence as:REα(P;Q) :=(∑x∈SuppQQ(x)∣∣∣P(x)Q(x) − 1∣∣∣α) 1α if α ≥ 1∆RE(P;Q) = maxx∈SuppQ∣∣∣P(x)Q(x) − 1∣∣∣ if α =∞The REα divergence generalizes several known metrics. RE1 is twice thestatistical distance ∆SD. In addition, RE∞ is exactly the relative error ∆RE.Finally, REαα = χα is an f -divergence for f(x) = |x− 1|α. The χα divergencewas first studied by Vajda [48], and χ2 (sometimes called Pearson Chi-squaredivergence) has recently been used in a cryptographic context [14].Proposition 1 provides a few properties of REα. Items 1 and 2 describe thebehavior of REα when α varies, and underline our point that REα is a continuoustrade-off between ∆SD and ∆RE. Items 3 and 4 show (with a tightness loss) thatREα and the Rényi divergence Rα are equivalent. Finally, Items 5 to 8 givecryptographically useful inequalities.Proposition 1. Let P,Q be two distributions over a countable space X, suchthat SuppP ⊂ SuppQ. The following properties hold:1. Monotonicity. α ≥ 1 7→ REα(P;Q) is a continuous non-decreasing func-tion.2. Upper bound from ∆SD and ∆RE. It holds that:REα(P;Q) ≤ ∆1/αSD (P;Q) ·∆1−1/αRE (P;Q).3. Upper bound on Rényi divergence. It holds that:Rα(P;Q) ≤ (1 + REα(P;Q))αα−1 . (3)4. Lower bound on Rényi divergence. If α is an even integer, then:REα(P;Q) ≤ 21−1/α(Rα(P;Q)α−1 − 1)1/α. (4)5. Multiplicative probability preservation. For any event E ⊂ X:P(E) ≤ Q(E)a−1a (1 + REα(P;Q)).106. Additive probability preservation. Let f : X → R be a function andp, q > 1 be such that 1/p+ 1/q = 1. It holds that:|E[f(P)]− E[f(Q)]| ≤ REp(P;Q) · E[|f(Q)|q]1/q.In particular, for any event E:|P(E)−Q(E)| ≤ REα(P;Q) · Q(E)1−1/α.7. Weak triangle inequality For distributions P,R,Q such that SuppP ⊂SuppR ⊂ SuppQ:∆RE(P,Q) ≤ ∆RE(P,R) +∆RE(R,Q) +∆RE(P,R) ·∆RE(R,Q)8. Data processing inequality. For any randomized function g, it holds that:REα(g(P), g(Q)) ≤ REα(P,Q).Most of the properties in Proposition 1 are proven in a rather straightforwardmanner by using either convexity, Minkowski’s inequality or generic properties off -divergences. This is no coincidence as the REα divergence is directly connectedto an f -divergence, and can also be interpreted as an Lα norm. The detailedproofs are given in Appendix D.Interestingly, the REα divergence does not have a native multiplicative prop-erty but the security analysis scales with the number of queries if we combineit with the Rényi divergence. However, the REα divergence benefits from nativeproperties that the Rényi divergence does not have, such as an additive prob-ability preservation property. This indicates that these divergences are comple-mentary to some extent: depending on the situation, one may be preferable tothe other.One last property allows the REα divergence to stand out. It was known forthe statistical distance but seems not to be known for other divergences. We callit proxy amplification and it will be discussed in Section 5, but for now we moveon to applications of the results of this section to prime number generators.4 Security proofs for Prime number generatorsWe now improve the security proofs for two known prime number generators:PRIMEINC [9] and Fouque-Tibouchi generator [18]. What we mean by securityis that the output distribution of the algorithm should be as close as possiblefrom the random one. For the first one, we show a bound on the Rényi divergencebetween the output distribution of the algorithm and the random distribution.For the second, we improve the results shown in [18] and show a bound onthe REα divergence between the output distribution and the random one. Wethen show how those proofs show security against known attacks towards primenumber generators.11Algorithm 1 PRIMEINC(x, s)Require: Parameters x, sEnsure: a prime number between x/2and x1: Sample odd p uniformly in Jx/2;xK2: for i = 1 to s do3: if p is prime then4: return p5: else6: p← p+ 27: end if8: end for9: return “failure”Algorithm 2 Fouque and Ti-bouchi’s prime number generatorRequire: x, ϵ, q ∝ x1−ϵEnsure: A prime number p ∈ J2;xK1: Sample a $←− (Z/qZ)∗2: repeat3: Sample t $←−{0, ...,⌊x−aq⌋}4: until p = a+ tq is prime5: return p4.1 Provable Security of the Prime Number Generator PRIMEINCThis section and Section 4.2 study prime number generators. The motivation forstudying them is given in Section 1.1. We first describe the simplest variant ofPRIMEINC in Algorithm 1.6Previous Analyses. Let P denote the output of PRIMEINC, and U the uniformdistribution over primes in Jx/2;xK. It was shown in [9] (resp. [18]) that underthe prime r-tuple conjecture:H1(P)H1(U)∼x→∞1 (5)∆SD(P;U) ≥ 0.86 + o(1) (6)In fact, we show in Appendix B a faster proof of (5) than the one from [9].On one hand, (5) seems to offer some security guarantees. However, Skórski [45,Corollary 4] showed that even when Shannon’s entropy H1(P) is close to itstrivial upper bound log2 |Supp(P)|, the collision entropy H2(P) might still below. Hence, (5) does not even preclude simple attacks such as common factorsattacks [25,30] on RSA. On the other hand, (6) is clearly a negative result.New Analysis. We now provide an improved security analysis. We first boundthe Rényi divergence between the output of PRIMEINC and the uniform dis-tribution. We then show that in common usecases, PRIMEINC may be safelyused without compromising security. We first recall the prime r-tuple conjecture,which is instrumental in the proofs of [9,18] and in ours.6 Security-efficiency trade-offs have been presented in [9], and OpenSSL implements avariant of PRIMEINC.12Conjecture 1 (Hardy-Littlewood’s prime r-tuple conjecture [24]).Let d = (d1, ..., dr) be a r-tuple of integers. Denote by πd(x) the number ofintegers n ≤ x such that for all i, n+ di is a prime number. It holds that:πd(x) ∼x→∞ Sd ·xlog(x)r,where Sd =∏p prime(pp−1)rp−νd(p)p−1 , and νd(p) is the number of distinct residueclasses modulo p of the tuple d.The following statement holds under the prime r-tuple conjecture.Theorem 1 (Theorem 1, [20]). Let x, k be integers. Then the number ofintegers n such that there are exactly k primes in Jn, n + λ log xK is ≃ x e−λλkk!when x→∞.In particular, if d(n) denotes the distance of n to the next larger prime, thenthe probability that d(n) > λ log x is e−λ.Theorem 2 is the main result of this section and implies that this algorithmis asymptotically secure by computing the Rényi divergence between its outputdistribution and the uniform distribution over a superset of its support.Theorem 2 (Security of PRIMEINC). Let P be the output distribution ofPRIMEINC with s = c log x and c > 0, and U be the uniform distribution overthe prime numbers between x/2 and x. Under the prime r-tuple conjecture:R∞(P;U) ≤ 2c(1 + oc,x(1)). (7)Equivalently, for all α ≥ 2,Hα(P) ≥ Hα(U)− log(2c) + oc,x(1). (8)The proof of Theorem 2 can be found in Appendix A. The proof computesan upper bound on the output probability of every prime and use Theorem 1 toconclude.Practical Implications I. Theorem 2 implies that for a constant number of calls toa prime number generator, a scheme secure with a uniform generator (ideal case)remains secure when using PRIMEINC instead (real case). One applicationis key generation for RSA signatures [44]. In the single-target setting, thereare two calls to the prime number generator.7 Combining (2) with (7), anyadversary breaking the real cryptosystem with probability ϵ will break the idealcryptosystem with probability at least ϵ/(2c + 1 + oc,x(1))2; therefore at most≈ 2 · log2(1 + 2c) bits of security are lost.7 This is true without loss of generality; even if more primes are generated and rejectedif they fail some requirements (e.g. being safe primes), the adversary only has accessto the product of exactly two outputs of the generator (p and q).13Practical Implications II. Even for a large number of queries, Theorem 2 providesecurity guarantees against some specific attacks. For example, taking α = 2in (8) gives a lower bound of n − log(2c) + oc,x(1) on the collision entropy ofPRIMEINC. As long as n ≥ λ+ log(2c) + oc,x(1) (which is always the case inpractice), this is more than enough to argue resistance against common factorsattacks [25,30].4.2 Provable Security of the Fouque-Tibouchi GeneratorWe now study a prime number generator proposed by Fouque and Tibouchi in[17,18]. Fouque and Tibouchi actually propose several algorithms, which providetrade-offs between simplicity and the number-theoretic conjectures they basetheir security on: the Friedlander-Granville-Montgomery conjecture, the gener-alized Riemann hypothesis or full unconditionality.We only study the simplest variant of Fouque and Tibouchi’s algorithms,and leave other variants for future work. The idea of this variant is to sample arandom number, and then resample only its most significant bits until a primenumber is found. It is described in Algorithm 2.Previous Analysis. In [18], it is shown under Conjecture 2 that:8∆SD(P;U)≪log xxϵ/4. (9)In addition, the average entropy consumption of Algorithm 2 is:(ϵ+ o(1)) · φ(q)q· (log x)2log 2. (10)The statistical distance requires ∆SD(P;U) ≤ 2−λ in order to provide λ bitsof security. As noted by [18], taking q to be a primordial (a product of smalldistinct primes) allows to reduce entropy consumption by a factor O(log log q).The Friedlander-Granville-Montgomery conjecture studies the quantity:π(x, q, a) := Card {p ≤ x prime | p ≡ a mod q} .The prime number theorem establishes that π(x, q, a) ∼x→∞xφ(q) log(x) when a, qare fixed. The Friedlander-Granville-Montgomery conjecture gives a bound onthe error given by this estimate.Conjecture 2. [Friedlander-Granville-Montgomery [19]] For q < x, (a, q) = 1 andall ϵ > 0, one has: ∣∣∣∣π(x, q, a)− π(x)φ(q)∣∣∣∣≪ϵ (xq)1/2xϵ/4.8 As stated in the preliminaries, this section will use Vinogradov’s notation, which iscommon in number theory: (f ≪s g)⇔ (f =s O(g)).14Since q ∼ x1−ϵ, it holds that(xq)1/2xϵ/4 ∼ x3ϵ/4, which is negligible com-pared to the ratio π(x)/φ(q)≫ xϵ/ log(x). Thus this conjecture implies:π(x, q, a) =π(x)φ(q)(1 + ox(1)) =xφ(q) · log x(1 + ox(1)).Theorem 3 is the main result of this section. It bounds the relative error andthe REα divergence between the output distribution of the Fouque-Tibouchigenerator and the uniform distribution over a superset of its support. Our resultis a generalization of Fouque-Tibouchi’s result as we show that the relative errorof any order decreases with the same exponential rate ϵ/4.Theorem 3 (Security of the Fouque-Tibouchi generator). Denote by Pthe output distribution of Algorithm 2 and by U the uniform distribution overJ2;xK. Under the Friedlander-Granville-Montgomery conjecture:∆RE(P;U| SuppP)≪log xxϵ/4and P[SuppPc] ≤ log(x)2x. (11)∀α ∈ (1,∞),REα(P;U)≪αlog xxϵ/4+(log xx2)1/α. (12)The proof can be found in Appendix C. It makes use of the Friedlander-Granville-Montgomery conjecture and of other classical arithmetical results.We now make explicit the implications of Theorem 3. Since the Fouque-Tibouchi generator is more provable security-oriented than PRIMEINC, it isunsurprising that Theorem 3 is intrinsically stronger than Theorem 2 and thatwe can assert security for more usecases than with PRIMEINC. Compared to[17,18], our analysis divides the entropy requirement by an order of magnitude.Practical Implications I. We study the generic security of RSA signature schemesin a multi-target setting. As an concrete example, consider a company producinga hardware security module (HSM); this HSM targets a bit-security λ = 128.At most 231 copies of it are produced, hence at most 232 queries to the primenumber generator are made. For RSA signatures, NIST and ENISA recommendRSA-3072. If HSMs generate RSA private keys with Algorithm 2, this translatesto x = 21536.A statistical distance argument requires – via (9) – that log xxϵ/4≤ 2−λ, whichhere is satisfied for ϵ ≥ 4λ log log xlog x ≈ 0.36. Using a Rényi-based argument in-stead, we combine Lemmas 3 and 4 with (11); this gives the milder requirementλQ (log x)22xϵ/2≤ 1, which is satisfied for ϵ ≥ 2 log(λQ) log log xlog x ≈ 0.076. Since entropyconsumption is essentially linear in ϵ, our new analysis provides a gain of afactor roughly 5. In practice, we take a primordial q =∏{p≤1049,p prime} p assuggested by [18]. As per (10), this gives an entropy consumption of less than20 000 uniform bits for generating two prime numbers.15Practical Implications II. We now study signature schemes based on the strongRSA assumption: this includes but is not limited to derivatives of Cramer-Shoupsignatures [12]. As in the previous example, we consider 128 bits of security,hence RSA-3072. Some of these schemes only require collision resistance of theprime number generator, in which case it is sufficient to use PRIMEINC sinceTheorem 2 showed that its collision entropy is high.Other schemes [26,50]only specify that the output of the generator shouldbe statistically close to the uniform distribution over primes in an interval, inwhich case it is more prudent to use Fouque and Tibouchi’s generator. Supposethat a user is queried at most 264 signatures. As before, an statistical distanceanalysis gives ϵ ≳ 0.36, whereas our Rényi-based analysis gives ϵ ≳ 0.12. Theentropy consumption is divided by 3.5 Proxy-Amplification and Application to Circuit-PrivateFHEIn this section, we come back to the REα divergence. We showed some standardproperties in Section 3 and now we focus on a quite unique property that wecall proxy amplification. In our opinion, this property justifies the definition ofour new divergence as other divergences do not enjoy a similar property – tothe best of our knowledge. We then apply this property to circuit-privacy onfully-homomorphic encryption in the fashion of [16].5.1 Proxy AmplificationProxy amplification is a unique property of the REα divergence. It generalizes anamplification property of the statistical distance [16, Lemma 2.3]. A major twistis that our property allows the REα divergence to “borrow” the amplificationof the statistical distance: REα will be made increasingly small even if it is > 1,as long as ∆SD < 1. For this reason, we call it proxy amplification.Proposition 2 (Proxy amplification). Let X be a finite space, f : X → X bea randomized function and α ∈ [1;+∞]. Suppose that for all x ∈ X, Supp f(x) =X and that there exists δ > 0 such that:∀a, b ∈ X, REα(f(a); f(b)) ≤ δ.Then, for all integer k ≥ 1:REα(fk(a); fk(b)) ≤ 2 · δ ·∆SD(f(a); f(b))k−1≤ δk/2k−2.The proof consists of using both convexity techniques and Minskowski’s in-equality along with the amplification property, already known for the statisticaldistance. It can be found in Appendix E. This property was already known in16the particular case of the statistical distance ∆SD [16, Lemma 2.3]. This gen-eralization is useful when the relative error is too big or infinite, which is forexample the case for shifted Gaussian distributions. Indeed, denote by DZ,σ,cthe Gaussian distribution of center x and standard deviation σ over Z. Then, inthe case of the relative error:∀x ̸= y,∆RE(DZ,σ,x;DZ,σ,y) =∞.On the other hand, REα(DZ,σ,x, DZ,σ,y) = eΘx,y,σ(α). Therefore, REα lies ina sweet spot between the statistical distance ∆SD and the relative error ∆RE.On one hand, ∆SD enjoys an amplification property, but not a multiplicativeprobability preservation. On the other hand, ∆RE enjoys both, but may not befinite. Finally, REα enjoys all three properties at once. Obviously in the caseof Gaussian distribution, a tailcut would be fine for security proofs but thisshows that from a theoretical point of view there are pathological cases with therelative error.5.2 Circuit-Private FHE: Setting the Washing Machine in EconomyModeWe recall that the motivation for circuit privacy in FHE schemes is given inSection 1.1. In the following section, we show how to get improved circuit privacyguarantees using a new analysis.The Ducas-Stehlé Strategy. Circuit privacy is guaranteed if the output distribu-tion of the ciphertext c′ does not depend on the circuit f . We briefly describe the“washing machine” strategy introduced by Ducas and Stehlé [16] to achieve it,and refer to [16] for a complete exposition. Ducas and Stehlé realize circuit pri-vacy by applying a randomized function Wash = Rerand◦Refresh that scramblesthe ciphertext c′. Here, Refresh is the bootstrapping operation, which reduces theciphertext noise down to some level. On the other hand, Rerand injects entropyin the ciphertext, see (13). [16] shows that applying Wash many times makesthe output distribution increasingly independent of the original ciphertext, andcompares this to a washing machine which repeats a cycle several times to “cleanup” the ciphertext.New Security Analysis. The security analysis in [16] leverages an amplificationproperty of the statistical distance. It guarantees λ bits of security if Wash isapplied O(λ) times. However, bootstrapping (Refresh) is extremely expensive,and it is desirable to perform it as rarely as possible. In this section, we providea new security analysis based on the amplification property of the REα diver-gence (Proposition 2). It allows us to claim that Wash needs only to be appliedO(log2 Q) times, where Q is the number of (homomorphically encrypted) queriesto f and the constants in O(·) are equivalent. Since one can expect log2 Q to bemuch smaller than λ, this entails much lighter requirements on the number ofcycles, hence a more practical protocol.17For brevity we only revisit [16, Section 4.1], but we expect our improvementsto be applicable to other examples in [16], as well as newer schemes such as [27].Suppose one wants to encrypt one bit µ ∈ {0, 1} under a private key s ∈ Znq withmodulus q and error rate less than η. The set of LWE ciphertexts decrypting toµ is:LWEqs(µ, η) ={(a, ⟨a, s⟩+ µ ·⌊q2⌋+ e)∈ Zn+1q∣∣∣ a ∈ Znq , |e| < ηq} .Correct decryption is ensured provided that η < 1/4. We now describe the Rerandfunction. Assume that the public key contains ℓ = O(n log q) encryptions of 0:∀i ≤ ℓ, ri = (ai, ⟨ai, s⟩+ ei) ∈ LWEqs(η, 0).For c ∈ LWEqs(η, µ), define:Rerand(pk, c) = c+∑iϵiri + (0, err), (13)where the ϵi’s are sampled uniformly in {0,±1} and err ∈ Zq is sampled fromthe discrete Gaussian DZ,σ,0 =: Dσ. We note that in [16], err is instead sampleduniformly over J−B,BK with a B suitable for correctness of decryption. In ourcase, the relative error precludes the use of this distribution as the support ofRerand(pk, c) should not depend on c. Therefore, every ciphertext should havea non-zero probability of appearance. We will set parameters that ensure thata bad ciphertext almost never (i.e with exponentially low probability) appears.To do so, we use Proposition 3.Proposition 3. For all k ∈ N, it holds that P[|t| > kσ; t←↩ Dσ] ≤ 2e−k22 .Next, Theorem 4 is the main result of this section and provides parametersthat guarantee that Ducas and Stehlé’s protocol is circuit-private. We boundthe relative error between the output distributions of Rerand applied on twodifferent ciphertexts. The theorem transfers directly to the function Wash =Rerand ◦ Refresh, as Refresh is deterministic and therefore has no impact on ouranalysis. This will subsequently allow to apply our amplification theorem onWash, followed by a Rényi divergence argument.Theorem 4. Given parameters η, n, q, s, ℓ. Let Q be the number of queries andk be the number of calls to the washing machine. Ifk >logQ2log(1/η4(ℓ+1))− log(8λ), (14)then there exists a standard deviation σ > 0 such that:1. The probability of Rerand outputting a bad ciphertext is < 2−λ.2. In the event that Rerand always output a good ciphertext, one has∀c1, c2 ∈ LWEqs(η, µ), ∆RE(Rerand(pk, c1);Rerand(pk, c2)) ≤ Q−1/2k.18The proof can be found in Appendix F. The outline goes like this. We find aT > 0 such that the probability of outputting an err of size > Tσ is lower than2−λ using Proposition 3. The correctness of decryption forces the condition(l + 1)ηq + Tσ < q/4.This gives an upper bound M on σ. For security, we require that the relativeerror ∆RE(Rerand(pk, c);Rerand(pk, c′) for two ciphertextst c, c′ is lower thanQ−1/2k in order to use the amplification property9 and Lemma 4. This will givea lower bound m on σ. To find a suitable σ, one only needs m ≤ M and thisgives the lower bound on k given in (14). Note that doing the same analysis withthe statistical distance leads to k > λlog( 1/η4(ℓ+1) ).Remark 1. Security is immediate via using Proposition 2 with taking X to bethe space LWEqs(µ, η) ={(a, ⟨a, s⟩+ µ · ⌊ q2⌋+ e ∈ Zn+1q∣∣ |e| < ηq}, and for therandom function f to be Wash (conditioned on the correctness of decryption).More precisely, we use the relative error by tailcutting the event where Washoutputs a bad ciphertext.Practical Implications. Consider a cloud service proposing homomorphic eval-uation of a proprietary function f over encrypted data. An adversarial user Atries to learn f or replicate it to some extent.To apply a Rényi-based argument, we need to formalize the problem that Atackles as a search problem. This is the most delicate part of our analysis andreally depends on the application: in the special case where the function f solvesa search problem, then one may simply say that the problem A tries to solve isto find a function f ′ which solves the same problem on a non-negligible fractionof the inputs. All usecases may not be formalized by a search problem, but webelieve that most practical ones can.If the conditions of Theorem 4 hold, then k applications of Wash will providea relative error ∆RE ≤ 1/√Q. We now apply the proof blueprint (a)⇒ (b)⇒ (c)outlined in Section 1.1. Combining Theorem 4 with Lemma 3 and applying (1)with α = Θ(λ), one can claim that O(log λ) bits of security are lost.10Applied on the washing machine technique of Ducas and Stehlé, our newanalysis allows to reduce the number of cycles by 2λlogQ for the same asymptoticbit security. For λ = 256 and Q = 264, this is an order of magnitude. Given thatbootstrapping often is a computational bottleneck, this can potentially make thewhole protocol faster by an order of magnitude.9 One would find it odd that we are not using the proxy amplification property herebut the computations we made showed that it wouldn’t give here a significantlybetter result than the amplification property for this application, so we chose not tocomplexify the computations done in the proof.10 Alternatively, one can replace Q by λQ in Theorem 4 and use Lemma 4; this resultsin a loss of O(1) bits of security and has a negligible effect on the parameters.19References1. Rohit Agrawal, Yi-Hsiu Chen, Thibaut Horel, and Salil P. Vadhan. Unifying com-putational entropies via kullback-leibler divergence. In Alexandra Boldyreva andDaniele Micciancio, editors, CRYPTO 2019, Part II, volume 11693 of LNCS, pages831–858. Springer, Heidelberg, August 2019.2. S. M. Ali and S. D. Silvey. A general class of coefficients of divergence of one dis-tribution from another. Journal of the Royal Statistical Society. Series B (Method-ological), 28(1):131–142, 1966.3. Shi Bai, Adeline Langlois, Tancrède Lepoint, Damien Stehlé, and Ron Steinfeld.Improved security proofs in lattice-based cryptography: Using the Rényi diver-gence rather than the statistical distance. In Tetsu Iwata and Jung Hee Cheon,editors, ASIACRYPT 2015, Part I, volume 9452 of LNCS, pages 3–24. Springer,Heidelberg, November / December 2015.4. Shi Bai, Tancrède Lepoint, Adeline Roux-Langlois, Amin Sakzad, Damien Stehlé,and Ron Steinfeld. Improved security proofs in lattice-based cryptography: Usingthe Rényi divergence rather than the statistical distance. Journal of Cryptology,31(2):610–640, April 2018.5. Daniel J. Bernstein, Yun-An Chang, Chen-Mou Cheng, Li-Ping Chou, NadiaHeninger, Tanja Lange, and Nicko van Someren. Factoring RSA keys from certifiedsmart cards: Coppersmith in the wild. In Kazue Sako and Palash Sarkar, editors,ASIACRYPT 2013, Part II, volume 8270 of LNCS, pages 341–360. Springer, Hei-delberg, December 2013.6. Olivier Binette. A note on reverse pinsker inequalities. IEEE Trans. InformationTheory, 65(7):4094–4096, 2019.7. Andrej Bogdanov, Siyao Guo, Daniel Masny, Silas Richelson, and Alon Rosen. Onthe hardness of learning with rounding over small modulus. In Eyal Kushilevitz andTal Malkin, editors, TCC 2016-A, Part I, volume 9562 of LNCS, pages 209–224.Springer, Heidelberg, January 2016.8. Florian Bourse, Rafaël del Pino, Michele Minelli, and Hoeteck Wee. FHE cir-cuit privacy almost for free. In Matthew Robshaw and Jonathan Katz, editors,CRYPTO 2016, Part II, volume 9815 of LNCS, pages 62–89. Springer, Heidelberg,August 2016.9. Jørgen Brandt and Ivan Damgård. On generation of probable primes by incremen-tal search. In Ernest F. Brickell, editor, CRYPTO’92, volume 740 of LNCS, pages358–370. Springer, Heidelberg, August 1993.10. Don Coppersmith. Finding a small root of a bivariate integer equation; factoringwith high bits known. In Maurer [33], pages 178–189.11. Don Coppersmith. Finding a small root of a univariate modular equation. InMaurer [33], pages 155–165.12. Ronald Cramer and Victor Shoup. Signature schemes based on the strong RSAassumption. In Juzar Motiwalla and Gene Tsudik, editors, ACM CCS 99, pages46–51. ACM Press, November 1999.13. Imre Csiszár. Eine informationstheoretische ungleichung und ihre anwendung aufden beweis der ergodizitat von markoffschen ketten. Magyar. Tud. Akad. Mat.Kutató Int. Közl, 8:85–108, 1963.14. Wei Dai, Viet Tung Hoang, and Stefano Tessaro. Information-theoretic indistin-guishability via the chi-squared method. In Jonathan Katz and Hovav Shacham,editors, CRYPTO 2017, Part III, volume 10403 of LNCS, pages 497–523. Springer,Heidelberg, August 2017.2015. Léo Ducas, Vadim Lyubashevsky, and Thomas Prest. Efficient identity-based en-cryption over NTRU lattices. In Palash Sarkar and Tetsu Iwata, editors, ASI-ACRYPT 2014, Part II, volume 8874 of LNCS, pages 22–41. Springer, Heidelberg,December 2014.16. Léo Ducas and Damien Stehlé. Sanitization of FHE ciphertexts. In Marc Fischlinand Jean-Sébastien Coron, editors, EUROCRYPT 2016, Part I, volume 9665 ofLNCS, pages 294–310. Springer, Heidelberg, May 2016.17. Pierre-Alain Fouque and Mehdi Tibouchi. Close to uniform prime number genera-tion with fewer random bits. In Javier Esparza, Pierre Fraigniaud, Thore Husfeldt,and Elias Koutsoupias, editors, ICALP 2014, Part I, volume 8572 of LNCS, pages991–1002. Springer, Heidelberg, July 2014.18. Pierre-Alain Fouque and Mehdi Tibouchi. Close to uniform prime number gener-ation with fewer random bits. IEEE Trans. Information Theory, 65(2):1307–1317,2019.19. John Friedlander and Andrew Granville. Limitations to the equi-distribution ofprimes i. Annals of Mathematics, 129(2):363–382, 1989.20. P. X. Gallagher. On the distribution of primes in short intervals. Mathematika,23(1):49, 1976.21. Craig Gentry. Fully homomorphic encryption using ideal lattices. In MichaelMitzenmacher, editor, 41st ACM STOC, pages 169–178. ACM Press, May / June2009.22. Sebastien Gerchinovitz, Pierre Ménard, and Gilles Stoltz. Fano’s inequality forrandom variables, 2017. https://arxiv.org/abs/1702.05985.23. Benedikt Gierlichs, Lejla Batina, Pim Tuyls, and Bart Preneel. Mutual informationanalysis. In Elisabeth Oswald and Pankaj Rohatgi, editors, CHES 2008, volume5154 of LNCS, pages 426–442. Springer, Heidelberg, August 2008.24. G. H. Hardy and J. E. Littlewood. Some problems of partitio numerorum; iii: Onthe expression of a number as a sum of primes. Acta Math., 44:1–70, 1923.25. Nadia Heninger, Zakir Durumeric, Eric Wustrow, and J. Alex Halderman. Min-ing your ps and qs: Detection of widespread weak keys in network devices. InTadayoshi Kohno, editor, USENIX Security 2012, pages 205–220. USENIX Asso-ciation, August 2012.26. Dennis Hofheinz, Tibor Jager, and Eike Kiltz. Short signatures from weaker as-sumptions. In Dong Hoon Lee and Xiaoyun Wang, editors, ASIACRYPT 2011,volume 7073 of LNCS, pages 647–666. Springer, Heidelberg, December 2011.27. Antoine Joux. Fully homomorphic encryption modulo fermat numbers. CryptologyePrint Archive, Report 2019/187, 2019. https://eprint.iacr.org/2019/187.28. Adeline Langlois, Damien Stehlé, and Ron Steinfeld. GGHLite: More efficientmultilinear maps from ideal lattices. In Phong Q. Nguyen and Elisabeth Oswald,editors, EUROCRYPT 2014, volume 8441 of LNCS, pages 239–256. Springer, Hei-delberg, May 2014.29. Arjen K. Lenstra, James P. Hughes, Maxime Augier, Joppe W. Bos, ThorstenKleinjung, and Christophe Wachter. Public keys. In Reihaneh Safavi-Naini andRan Canetti, editors, CRYPTO 2012, volume 7417 of LNCS, pages 626–642.Springer, Heidelberg, August 2012.30. Arjen K. Lenstra, James P. Hughes, Maxime Augier, Joppe W. Bos, ThorstenKleinjung, and Christophe Wachter. Ron was wrong, whit is right. CryptologyePrint Archive, Report 2012/064, 2012. http://eprint.iacr.org/2012/064.31. San Ling, Duong Hieu Phan, Damien Stehlé, and Ron Steinfeld. Hardness of k-LWE and applications in traitor tracing. In Juan A. Garay and Rosario Gennaro,21editors, CRYPTO 2014, Part I, volume 8616 of LNCS, pages 315–334. Springer,Heidelberg, August 2014.32. Takahiro Matsuda, Kenta Takahashi, Takao Murakami, and Goichiro Hanaoka.Improved security evaluation techniques for imperfect randomness from arbitrarydistributions. In Dongdai Lin and Kazue Sako, editors, PKC 2019, Part I, volume11442 of LNCS, pages 549–580. Springer, Heidelberg, April 2019.33. Ueli M. Maurer, editor. EUROCRYPT’96, volume 1070 of LNCS. Springer, Hei-delberg, May 1996.34. Daniele Micciancio and Michael Walter. Gaussian sampling over the integers: Ef-ficient, generic, constant-time. In Jonathan Katz and Hovav Shacham, editors,CRYPTO 2017, Part II, volume 10402 of LNCS, pages 455–485. Springer, Heidel-berg, August 2017.35. Daniele Micciancio and Michael Walter. On the bit security of cryptographic prim-itives. In Jesper Buus Nielsen and Vincent Rijmen, editors, EUROCRYPT 2018,Part I, volume 10820 of LNCS, pages 3–28. Springer, Heidelberg, April / May2018.36. Preda Mihailescu. Fast generation of provable primes using search in arithmeticprogressions. In Yvo Desmedt, editor, CRYPTO’94, volume 839 of LNCS, pages282–293. Springer, Heidelberg, August 1994.37. Ilya Mironov. Renyi differential privacy. Proceedings of 30th IEEE ComputerSecurity Foundations Symposium, 2017. http://arxiv.org/abs/1702.07476.38. Tetsuzo Morimoto. Markov processes and the h-theorem. Journal of the PhysicalSociety of Japan, 18(3):328–331, 1963.39. Matús Nemec, Marek Sýs, Petr Svenda, Dusan Klinec, and Vashek Matyas. Thereturn of coppersmith’s attack: Practical factorization of widely used RSA moduli.In Bhavani M. Thuraisingham, David Evans, Tal Malkin, and Dongyan Xu, editors,ACM CCS 2017, pages 1631–1648. ACM Press, October / November 2017.40. Thomas Pöppelmann, Léo Ducas, and Tim Güneysu. Enhanced lattice-based sig-natures on reconfigurable hardware. In Lejla Batina and Matthew Robshaw, ed-itors, CHES 2014, volume 8731 of LNCS, pages 353–370. Springer, Heidelberg,September 2014.41. Thomas Prest. Sharper bounds in lattice-based cryptography using the Rényidivergence. In Tsuyoshi Takagi and Thomas Peyrin, editors, ASIACRYPT 2017,Part I, volume 10624 of LNCS, pages 347–374. Springer, Heidelberg, December2017.42. Thomas Prest, Dahmun Goudarzi, Ange Martinelli, and Alain Passelègue. Unifyingleakage models on a Rényi day. In Alexandra Boldyreva and Daniele Micciancio,editors, CRYPTO 2019, Part I, volume 11692 of LNCS, pages 683–712. Springer,Heidelberg, August 2019.43. Oded Regev. On lattices, learning with errors, random linear codes, and cryptog-raphy. In Harold N. Gabow and Ronald Fagin, editors, 37th ACM STOC, pages84–93. ACM Press, May 2005.44. Ronald L. Rivest, Adi Shamir, and Leonard M. Adleman. A method for obtainingdigital signatures and public-key cryptosystems. Communications of the Associa-tion for Computing Machinery, 21(2):120–126, 1978.45. Maciej Skórski. Shannon entropy versus Renyi entropy from a cryptographic view-point. In Jens Groth, editor, 15th IMA International Conference on Cryptographyand Coding, volume 9496 of LNCS, pages 257–274. Springer, Heidelberg, December2015.2246. John Steinberger. Improved security bounds for key-alternating ciphers viahellinger distance. Cryptology ePrint Archive, Report 2012/481, 2012. http://eprint.iacr.org/2012/481.47. Katsuyuki Takashima and Atsushi Takayasu. Tighter security for efficient lat-tice cryptography via the Rényi divergence of optimized orders. In Man Ho Auand Atsuko Miyaji, editors, ProvSec 2015, volume 9451 of LNCS, pages 412–431.Springer, Heidelberg, November 2015.48. Igor Vajda. χα-divergence and generalized fisher information. In Transactions ofthe Sixth Prague Conference on Information Theory, Statistical Decision Functionsand Random Processes, page 223. Academia, 1973.49. Tim van Erven and Peter Harremoës. Rényi divergence and kullback-leibler diver-gence. IEEE Trans. Information Theory, 60(7):3797–3820, 2014.50. Shota Yamada, Goichiro Hanaoka, and Noboru Kunihiro. Two-dimensional rep-resentation of cover free families and its applications: Short signatures and more.In Orr Dunkelman, editor, CT-RSA 2012, volume 7178 of LNCS, pages 260–277.Springer, Heidelberg, February / March 2012.23A Proof of Theorem 2Proof. Denote by N the number of integers between x/2 and x that do not leadto failure, an integer n0 ∈ Jx/2, xK doest not lead to failure if it is odd andd(n) < 2c log x, by Theorem 1 one has N ≃ x4 (1 − e−2c). For a prime p denoteby d(p) its distance to the first prime lower than him. Then, the algorithmwill output p if and only if n0 ∈Kp− d(p), pK if d(p) ≤ 2c log x and if and only ifn0 ∈Kp−2c log x, pK otherwise. Therefore, the probability P(p) that the algorithmoutputs p isP(p) =d(p)2N, if d(p) ≤ 2c log x.2c log x2N, if d(p) > 2c log x.Either way, one has P(p) ≤ c log xN . We now compute the Renyi divergence be-tween P and U .P(p)U(p)≤(π(x)− π(x2 ))c log xN≤ 2c1− e−2c(1 + ox(1)).The inequality on the entropies comes from the relation Hα(P) −Hα(U) =− log2 Rα(P;U).B Proof of quasi-optimal entropy of the outputdistribution of PRIMEINCProof. Let c > 0 and P be the output distribution of PRIMEINC with s =c log x for a given integer x. Let N be the number of integers in Jx/2, xK thatdo not lead to failure. As in appendix A, for all prime p ∈ Jx/2, xK, one hasP(x) ≤ c log xN and N ≃x4 (1− e−2c). Now, H1(U) ≃ log(x2 log x)and H1(P)H1(U) ≤ 1because the map x > 0 7→ log 1x is convex. We bound H1(P) from below, thesummations are over p prime.H1(P) =∑p∈Jx/2,xKP(p) log1P(p)≥ log(Nc log x)≥(log(x2 log x)− log 2c)(1 + ox(1)).Dividing by H1(U), one gets1 ≥ H1(P)H1(U)≥1− log 2clog(x2 log x) (1 + ox(1)) .And this concludes the proof.24C Proof of Theorem 3Proof. We prove both items separately:1. If p is a prime such that p = a+ tq, then:P[P = p] = 1φ(q)π(x, q, a).Therefore:|P[P = p]− P[U = p]|P[U = p]=1π(x, q, a)∣∣∣∣π(x)φ(q) − π(x, q, a)∣∣∣∣≪ log xxϵx3ϵ/4(1 + ox(1))≪ log x · x−ϵ/4.If p is a prime divisor of q then P[X = p] = 0 and thus |P[P=p]−P[U=p]|P[U=p] = 1,however there are w(q) ≤ log x such prime numbers and their probability ofappearance is w(q)/π(x) ≤ log2(x)x , which bounds P[SuppPc]. This concludesthe proof.2. We compute now the REα divergence between the two distributions.REα(P;U)= ∑a∈(Z/qZ)∗∑a+tq≤x, prime∣∣∣ 1φ(q)π(x,q,a) − 1π(x) ∣∣∣α1π(x)α−1+∑p|q,p prime1π(x)1/α= ∑a∈(Z/qZ)∗1π(x)π(x, q, a)α−1∣∣∣∣π(x)φ(q) − π(x, q, a)∣∣∣∣α + w(q)π(x)1/α≪(φ(q)απ(x)αx3αϵ/4 +log(x)π(x))1/αSince for all x, y ≥ 0, (x+ y)1/α ≤ x1/α + y1/α, one has≪αφ(q)x3ϵ/4π(x)+(log(x)2x)1/α≪αlog(x)xϵ/4+(log(x)2x)1/α.This concludes the proof.D Proof of Proposition 1Proof. We prove each item separately.251. This is done using concavity and Jensen’s inequality. Take β > α > 0, thenthe map x > 0 7→ xα/β is concave. Therefore:REα(P;Q) =∑x∈XQ(x)[∣∣∣∣P(x)−Q(x)Q(x)∣∣∣∣β]α/β1/α≤[∑x∈XQ(x)∣∣∣∣P(x)−Q(x)Q(x)∣∣∣∣β]α/β1/α = REβ(P;Q).2. The result is an immediate application of [6, Theorem 1].3. We recall Minkowski’s inequality:∀p ≥ 1,(∑k|xk + yk|p)1/p≤(∑kxpk)1/p+(∑kypk)1/p.We now prove Item 3:Rα(P;Q) =(∑xQ(x)(P(x)Q(x))α)1/α−1≤{∑x(Q(x)1/α∣∣∣∣P(x)Q(x) − 1∣∣∣∣+Q(x)1/α)α}1/α αα−1≤(∑xQ(x)∣∣∣∣P(x)Q(x) − 1∣∣∣∣α)1/α+(∑xQ(x))1/α αα−1 (15)≤ (1 + REα(P;Q))αα−1Minkowski’s inequality is used in (15).264. Remember that α is even here. Therefore, for all x, y ∈ R, |x+ y|α = (x +y)α =∑k=0(αk)xkyα−k. Using this, we getREα(P;Q) = ∑x∈SuppQQ(x)∣∣∣∣P(x)Q(x) − 1∣∣∣∣α1/α=(∑xQ(x)α∑k=0(αk)(P(x)Q(x))k(−1)α−k)1/α= α∑k=0,k even(αk)Rk(P;Q)k−1︸︷︷︸≤Rα(P;Q)α−1−α∑k=0,k odd(αk)Rk(P;Q)k−1︸︷︷︸≥11/α≤Rα(P;Q)α−1α∑k=0,k even(αk)︸︷︷︸=2α−1−α∑k=0,k odd(αk)︸︷︷︸=2α−11/α= 21−1/α(Rα(P;Q)α−1 − 1)1/α5. This is a direct corollary of (1) and Item 3.6. Again, we use Minkowski’s inequality:|E[f(P)]− E[f(Q)]| =∣∣∣∣∣∑x∈Xf(x) · |P(x)−Q(x)|∣∣∣∣∣≤∑x∈X|f(x)| · |Q(x)|1/q |P(x)−Q(x)||Q(x)|1/q≤(∑x∈XQ(x) |f(x)|q)1/q·(∑x∈X|P(x)−Q(x)|pQ(x)p−1)1/p= REp(P;Q) · E[|f(Q)|q]1/q.7. This is a simple use of the triangle inequality. Let x ∈ X,∣∣∣∣P(x)Q(x) − 1∣∣∣∣ = ∣∣∣∣P(x)R(x) · R(x)Q(x) − 1∣∣∣∣=∣∣∣∣(P(x)R(x) − 1)(R(x)Q(x)− 1)+(P(x)R(x)− 1)+(R(x)Q(x)− 1)∣∣∣∣≤ ∆RE(P;R) +∆RE(R;Q) +∆RE(P;R) ·∆RE(R;Q)8. REαα is an f -divergence for f(x) = |x− 1|α. The data processing inequalityis true for all f -divergences by Lemma 1, hence the result follows.27E Proof of Proposion 2Proof. We first prove this for α <∞. We use the notation [f(t) = c] := P[f(t) =c]. We prove this by induction on k. This is true for k = 1, suppose the propo-sition to be true for an integer k and take a, b ∈ X, thenREα(fk+1(a); fk+1(b)) =(∑x∈X∣∣[fk+1(a) = x]− [fk+1(b) = x]∣∣α[fk+1(b) = x]α−1)1/α=(∑x∈X∣∣[fk+1(a) = x]− [fk+1(b) = x]∣∣α(∑z∈X [fk(b) = z][f(z) = x])α−1)1/αSince α ≥ 1, the map x > 0 7→ x−(α−1) is convex, therefore:REα(fk+1(a); fk+1(b)) ≤( ∑x,z∈X[fk(b) = z]∣∣[fk+1(a) = x]− [fk+1(b) = x]∣∣α([f(z) = x])α−1)1/α(16)We now bound the numerators in (16):∣∣[fk+1(a) = x]− [fk+1(b) = x]∣∣=∣∣∣∣∣∣∑y∈X[fk(a) = y][f(y) = x]− [fk(b) = y][f(y) = x]∣∣∣∣∣∣ (17)=∣∣∣∣∣∣∑y∈X([fk(a) = y]− [fk(b) = y])([f(y) = x]− [f(z) = x])∣∣∣∣∣∣ (18)≤∑y∈X∣∣[fk(a) = y]− [fk(b) = y]∣∣ |[f(y) = x]− [f(z) = x]| (19)28Since∑y∈X([fk(a) = y] − [fk(b) = y])[f(z) = x] = 0, (17) and (18) are equal.Combining (16) and (19) yields (all the summations are over X):REα(fk+1(a); fk+1(b))≤∑x,z[fk(b) = z](∑y∣∣[fk(a) = y]− [fk(b) = y]∣∣ |[f(y) = x]− [f(z) = x]|)α([f(z) = x])α−11/α(20)≤∑y(∑x,z[fk(b) = z](∣∣[fk(a) = y]− [fk(b) = y]∣∣ |[f(y) = x]− [f(z) = x]|)α([f(z) = x])α−1)1/α(21)≤∑y(∑z[fk(b) = z] ·REα(f(y); f(z))α ·∣∣∣[fk(a) = y]− [fk(b) = y]∣∣∣α)1/α≤δ∑y∣∣∣[fk(a) = y]− [fk(b) = y]∣∣∣ = δ · 2 ·∆SD(fk(a); fk(b))(22)Finally, we use the amplification property with ∆SD, hence ∆SD(fk(a), fk(b)) ≤∆SD(f(a), f(b))k (see [16]) and that ∆SD = 12RE1 to getREα(fk+1(a); fk+1(b)) ≤ δk+1/2k−1Equation (20) is a sum of the form(∑x,z∣∣∣∑y gy(x, z)∣∣∣α)1/α = ∣∣∣∣∣∣∑y gy∣∣∣∣∣∣α,where ||·||α denotes the Lα norm. By Minkowski’s inequality, it is upper boundedby: ∑y(∑x,z|fy(x, z)|α)1/α=∑y||fy||α ,giving (21). The result follows. By a continuity argument, it also follows forα =∞.F Proof of Theorem 4Proof. Denote c′ = c+∑ϵiri = (a′, ⟨a′, s⟩+ µ⌊q/2⌋+ e′). By the leftover hashlemma [43, Section 5], the distribution of a′ is exponentially close to the uniformdistribution. So the only leakage coming from c′ is from e′ and one has |e′| <(ℓ + 1)ηq. Let λ denote the level of security required, then one wants to find Tsuch that P[|t| > Tσ; t ←↩ Dσ] < 2−λ. Using the inequality from Proposition 3,we take T =√2λ.This means that with probability > 1 − 2−λ, Rerand(pk, c) = (a, ⟨a, s⟩ +µ⌊q/2⌋+e) will have an error such that |e| < (ℓ+1)ηq+Tσ. Therefore, correctnessof decryption is ensured if:(l + 1)ηq + Tσ < q/4. (23)29Let us compute the relative error in the setup where Rerand always gives agood ciphertext. Let x, y ∈ [−(ℓ+ 1)ηq, (ℓ+ 1)ηq].∆RE(DZ;σ,x, DZ,σ,y) = max|t|<(ℓ+1)ηq+Tσ∣∣∣∣P[DZ,σ,x = t]P[DZ,σ,y = t] − 1∣∣∣∣= max∣∣∣∣e (t−y)2σ2 − (t−x)2σ2 − 1∣∣∣∣= max∣∣∣e (2t−(x+y))σ y−xσ − 1∣∣∣≤ max(e2(ℓ+1)ηq·(4(ℓ+1)ηq+2Tσ)σ2 − 1, 1− e−2(ℓ+1)ηq·(4(l+1)ηq+2Tσ)σ2 )For the sake of clarity, we note β = 2(ℓ+1)ηq and u = β(2β+2Tσ)σ2 . It holds that:∆RE(DZ,σ,x;DZ,σ,y) ≤ max(eu − 1, 1− e−u).Our goal is have ∆RE ≤ Q−1/2k with k the number of iterations and Q beingthe number of queries. It holds that:eu − 1 ≤ Q−1/2k ⇒ u ≤ log(1 +Q−1/2k)1− e−u ≤ Q−1/2k ⇒ u ≤ umax,Where umax = log(11−Q−1/2k)≥ log(1 +Q−1/2k). This yields:σ2 − 2Tβumaxσ − 2 β2umax≥ 0 (24)Solving this polynomial inequality of degree 2 gives:σ ≥ Tβumax(1 +√1 +2umaxT 2)∼ 2TβQ1/2k (25)Combining (23) and (25) one has:2TβQ1/2k < σ logQ2log(1/η16T 2(ℓ+1))30Finally, we replace T =√2λ to getk >logQ2log(1/η32λ(ℓ+1)) = logQ2log(1/η4(ℓ+1))− log(8λ)

than

Abstract

Extracted data

than

Abstract

Extracted data

Related items

Related items