fccjxxw.com

非常超级学习网 学习超级帮手

非常超级学习网 学习超级帮手

当前位置：首页 >> >> # Model-Based Testing of Cryptographic Protocols

Model–based Testing of Cryptographic Protocols

Dean Rosenzweig1,2 , Davor Runje2 , and Wolfram Schulte1

Microsoft Research schulte@microsoft.com 2 University of Zagreb dean@math.hr, davor.runje@fsb.hr

1

Abstract. Modeling is a popular way of representing the behavior of a system. A very useful type of model in computing is an abstract state machine which describes transitions over ?rst order structures. The general purpose model-based testing tool SpecExplorer (used within Microsoft, also available externally) uses such a model, written in AsmL or Spec#, to perform a search that checks that all reachable states of the model are safe, and also to check conformance of an arbitrary .NET implementation to the model. Spec Explorer provides a variety of ways to cut down the state space of the model, for instance by ?nitizing parameter domains or by providing predicate abstraction. It has already found subtle bugs in production software. First order structures and abstract state machines over them are also a useful way to think about cryptographic protocols, since models formulated in these terms arise by natural abstraction from computational cryptography. In this paper we explain this abstraction process, ‘experiments as structures’, and argue for its faithfulness. We show how the Dolev–Yao intruder model ?ts into SpecExplorer. In a word, the actions of the Dolev– Yao intruder are the ‘controllable’ actions of the testing framework, whereas the actions of protocol participants are the ‘observable’ actions of the model. The unsafe states are the states violating say Lowe’s security guarantees. Under this view, the general purpose software testing tool quickly ?nds known attacks, such as Lowe’s attack on the Needham– Schroeder protocol.

Introduction: Why Yet Another Formal Model

A new ‘behavioral’ theory of algorithms has been developed in recent years in a series of papers by Y.Gurevich, A.Blass [Gur00,BG03,BG04a,BG04b,Gur05], and also B.Rossman and the authors [RR05]. The gist is that algorithms can be mathematically captured at their own native level of abstraction - ex. the native level of abstraction of the Euclidean algorithm is that of Euclidean rings. Algorithms operate over abstract ?rst-order structures, well studied and familiar in mathematical logic, algebra and abstract mathematics in general. The techniques developed for behavioral theory suggest a natural representation of Dolev-Yao assumptions in ?rst-order structures, and a natural mapping of ad-hoc notations present in abstract models of cryptography. Unlike the

static abstract models, which necessarily invoke additional proof-theoretic devices to capture dynamic aspects, the behavioral theory explicitly targets the dynamic behavior of algorithms semantically. By recent work on behavioral theory [BG04a,BG04b,RR05,Gur05], this also includes interactive algorithms talking to an environment between steps, and within a step, allowing us to represent the abstract content of oracle algorithms and adversary games typical of computational cryptography directly. In the framework of intra-step interactive algorithms exact abstract representations of computational security notions, de?ned in terms of adversary games, emerge clearly. The experiments of asymptotic computational cryptography can be naturally represented in terms of interactive algorithms over ?rst-order structures, this is our experiments-as-structures paradigm, providing a setting for soundness/completeness proofs. The abstract content of these proofs gets more clearly separated from the probabilistic aspects. In this paper we execute a small initial segment of this program, in case of confusion-free asymmetric encryption. Abstract models for the standard asymptotic security notions in this case are provided, with proofs of their soundness (under the assumption of acyclicity) and completeness. The relation of these proofs to proofs in the literature [AR02,MW04a,AJ01,Ban04,ABS05] can best be described as extraction of abstract content. We also brie?y indicate how the assumptions of confusion-freeness and acyclicity can be relaxed in our setting. Section 1 is a (necessarily cursory) overview of the behavioral theory of algorithms, essentially referring the reader to the literature. Section 2 is a brief summary of the relevant assumptions of asymptotic computational cryptography in the asymmetric (public key) case. Section 3 presents the experiments-asstructures paradigm and our abstract model of cryptographic adversary games. Section 4 contains sketches of soundness and completeness proofs, and how the Abadi-Rogaway expression language variant embeds into our framework. Testing model for public key protocols is in Section 5, together with an example of rediscovery of Lowe’s attack on the Needham–Schroeder protocol by SpecExplorer. In addition to quoted cryptographic literature, some understanding of the framework as presented in [RR05] is expected of the reader.

1

Behavioral Theory of Algorithms

The behavioral theory of algorithms is not an attempt to question the ChurchTuring thesis, saying that every computable function over natural numbers can be computed by a Turing machine, or the stronger implicit thesis, actually argued for by Turing, that every algorithm can be simulated by a Turing machine. The aim of the behavioral theory is to make semantical distinctions ?ner than that precise. While algorithms get implemented (simulated) exclusively over bits these days, they are often intended to operate over much more abstract objects, abstract data-structures of algebraic or geometric or analytic or even not explicitly mathematical character. The behavioral theory aims to capture algorithms as they are intended, at their own level of abstraction. 2

The requirement of “capturing algorithms at their own level of abstraction” is made precise as the requirement of simulation step-by-step. The technology to achieve this is using ?rst-order structures, well known to capture faithfully arbitrary static mathematical situations, as states of algorithms. The dynamics, the step, is also de?ned in terms of the abstract state. This philosophy leads to a sharp mathematical de?nition, technically developed in [Gur00] and overviewed in [BG03], computationally realized in the theoretical programming language of Abstract State Machines [Gur00] and the implemented programming languages AsmL [AsmL] and Spec# [Spec#]. Models written in these modelling languages are used by a model-based software testing tool SpecExplorer, also developed at Microsoft Research [SpecExp]. 1.1 Interactive Algorithms

Interactive algorithms issue queries to the environment, which contain labels and data, and receive replies, which are data, elements of algorithm’s state, within a step. This mechanism allows a clean separation of computational (the algorithm) and declarative (the environment) aspects, and naturally models nondeterminism, function calls, interaction with oracles, input and output,. . . The full theory of (ordinary) interactive algorithms is developed in [BG04a,BG04b]; overviews are given in [Gur05] and [RR05]. All algorithms in this paper are assumed to be small-step ordinary interactive algorithms in the sense of [BG04a,BG04b,Gur05]. 1.2 Accessibility, Reachability and Indistinguishability

The notions of accessibility of objects, reachability and indistinguishability of states, as introduced in [RR05], will be important here. An object is accessible at a state if it is the value of a term there. A state Y is reachable from a state X if there is an algorithm turning X to Y . Two states X, Y are distinguishable if there is an algorithm turning them into states distinct by values of a speci?c term. Structures X, Y of the same vocabulary are similar, written X ? Y if they induce the same equivalence on ground terms: V al(t1 , X) = V al(t2 , X) i? V al(t1 , Y ) = V al(t2 , Y ) Precise de?nitions and the theory behind these notions can be found in [RR05]. Here we shall repeatedly use the following results from [RR05](where Y ?X is the set of di?erences of two states over the same carrier, see [Gur00,BG04a,RR05] for de?nitions): Theorem 1. State Y is reachable from state X i? – X and Y have the same base set; and – Y ? X is ?nite and every element in Y ? X is accessible. Theorem 2. States X and Y are indistinguishable by small–step algorithms i? X ?Y. 3

1.3

Background Structures and Importing/Creating

An algorithms often needs to create a new object. A Turing machine often needs to access a new tape location never used before. In the TM case it obviously doesn’t matter whether we conceive its tape as ?nite, creating new locations as needed, or as in?nite, with all locations possibly needed given in advance. In the latter case locations get activated as the TM visits them for the ?rst time. The case of a ?rst-order structure is the same, a reserve pool (“the heap”) of su?ciently many fresh amorphous objects can be given in advance, to be accessed as needed. For interactive algorithms, they are available to the environment to be returned in reply to an appropriate query (get me a new . . . ). The reserve elements are amorphous in the sense that no “signi?cant” functions are de?ned on them, or denote them as values. For abstract cryptography the amorphous reserve objects will represent random coins. But if we have some infrastructure de?ned on all objects, such as ordered pairs and/or ?nite sets and/or encryptions, it would be both unnatural and very boring to have to establish all the infrastructure over a new element each time one is introduced, brought forward from the reserve. The notions of background structure and background class [BG00] serve exactly this purpose: the axioms for a background class of [BG00] specify what kind of structure can exist over amorphous atoms without imposing any speci?c properties on them except for identity. See [BG00,RR05] for de?nitions of background classes, background of algorithms, exposed elements, active part, reserve. A structure X is explicitly atom–generated if the smallest substructure of X that includes all atoms is X itself. All background structures in the paper are assumed to be explicitly atom-generated. Atomic support of a set S of elements of a structure X from a background class K is the set of atoms of the envelope of S, the smallest K-substructure containing S. Corollary 1. If the atomic support SupX ({x}) of an element x is accessible in a state X, then x is accessible in X. We assume that the set of exposed elements is ?nite, but not necessary uniformly bounded, in every state. Remember that the foreground of an algorithm is its (generalized) memory, storing input data and results of previous calculations. As such, after a ?nite number of algorithm steps, only a ?nite number of locations can be changed. Let 0X denote the reduct of X to the background vocabulary, the structure obtained by “forgetting” all foreground functions in state X. We assume that all states have an in?nite but countable reserve. It follows immediately from the axioms of [BG00] that if X and Y are K-states over the same carrier, then their background reducts are isomorphic 0X ? 0Y . = Theorem 3. Let X be a state with background BC. Then there is an algorithm A and an injective answer function α appropriate for 0X with only reserve elements in its codomain such that X = A(0X , α). 4

2

2.1

Computational Cryptography

Encryption Schemes

An asymmetric encryption scheme Π is a tuple of polytime algorithms (K, I, E, D) K : Parameter × Coins ?→ DecryptionKey I : DecryptionKey ?→ EncryptionKey E : EncryptionKey × String × Coins ?→ Ciphertext ∪ {⊥} D : DecryptionKey × String ?→ Plaintext ∪ {⊥} where String denotes the set of ?nite strings over {0, 1}, domains EncryptionKey, DecryptionKey, Ciphertext, Plaintext are subsets of String, ⊥ is a distinguished string representing failure of the algorithm, and Coins is the set of all in?nite strings over {0, 1} . The polytime assumption for K means time polynomial in η (not the size of its string representation) and ignores the Coins argument representing random coin ?ips. Suppressing the Coins argument K, E become probabilistic polytime algorithms, and K(η, c), E(k, m, c), D(K, m) are, according to tradition, often written as K(η), Ek (m), DK (m) respectively. The key-inversion algorithm I returns an encryption key matching the decryption key. Remark 1 (Usual Assumptions). We require that – DK (Ek (m)) = m whenever k = I(K), for every key K sampled from K(η) and every plaintext m such that Ek (m) doesn’t fail; – the Plaintext domain is the set of all m for which, for some EncryptionKey k, Ek (m) doesn’t fail; Ciphertext is the corresponding codomain; – if K, K are two outputs of K(η) for the same η, then ? K, K have the same length; ? k = I(K), k = I(K ) have the same length; ? if m, m are strings of the same length, then Ek (m) doesn’t fail if and only if Ek (m ) doesn’t fail, and then the encryptions have the same length. Remark 2. Syntax of an asymmetric encryption scheme is usually de?ned as a triple of algorithms (K, E, D), where K returns a pair of both encryption and decryption keys [BDPR98]. But then I is simply a projection and the decryption algorithm simply ignores one of the parameters. We ?nd our variant more convenient for the purpose of abstract modeling. We also assume a (polytime) encoding of ordered pairs, which means a triple of functions Σ = (P, F, S), where P is a binary pairing function on strings, and F and S are unary projections, with the usual properties. We also assume a type-?aw preventing tagging scheme, ensuring that the codomains of K, E, I, P are pairwise disjoint, and that neither of them contains ⊥. We also assume there is a polynomial time algorithm T distinguishing the codomains of functions K, I, E, P, and the codomain of the function N introduced bellow. 5

2.2

Notions of Security

Notions of security of encryption schemes are typically based on a notion of indistinguishability, represented by two sequences of oracles of the same length, G G F F the good-oracles O1 , . . . , On and the fake-oracles O1 , . . . , On . Each of the sequences gets initialized by randomly generating a sequence of keys to be used by respective oracles, good-init and fake-init. The oracles and the initializations are implicitly parameterized by the encryption scheme Π and possibly the pairing scheme Σ, but we shall drop this from the notation. Some data resulting from the initialization can be passed to the adversary algorithm as parameters—we consider this to be a part of the initialization. Let us call the initialization and oracle data just ATT, and let the notion of security de?ned by ATT be INDATT. The idea is that no PPT-limited adversary can distinguish whether she is working with the good or the fake oracles: De?nition 1. Let A be an algorithm working with n oracles. Its advantage for IND-ATT is Advind-att (A) = Pr[good-init : A(. . .)O1 ,...,On = 1] ? Π Pr[fake-init : A(. . .)O1 ,...,On = 1] The encryption scheme Π is IND-ATT secure if no probabilistic polytime algorithm A can guess which set of oracles it is provided with, with probability negligible in the security parameter η: Advind-att (A) is negligible. Π The A(. . .) notation denotes the adversary algorithm called with any parameters that the initialization chooses to provide. Thus the notion of security is completely characterized by the initializations and the oracles selected. By negligible we mean, throughout this paper, polynomially negligible func1 tions: f (n) such that for every c for all su?ciently large n we have f (n) ≤ nc , and by overwhelming those negligibly close to 1. We de?ne oracles characterizing notions of securities called indistinguishability under chosen–plaintext attack and indistinguishability under adaptive chosen– ciphertext attack, denoted with IND-CPA and IND-CCA, respectively. Example 1 (IND-CPA). – Let good-init be K ← K(η), passing along to the adversary algorithm k = I(K). – Let good-oracles be O with O(m1 , m2 ) = Ek (m1 ). – Let fake-init be as good-init. – Let fake-oracles be O with O(m1 , m2 ) = Ek (m2 ), where k = I(K). This de?nes the notion of security known as IND-CPA, “security under known plaintext attack”. Example 2 (IND-CCA). – Let good-init, fake-init be as for IND-CPA. 6

F F G G

– Let good-oracles be O, Od where O is as good-oracles of IND-CPA, and Od (e) = DK (e) given that e is not an output obtained from O; if it is, then Od (e) fails. – Let fake-oracles be O, Od , with O as fake-oracle of IND-CPA and Od as in good-init. This de?nes a strictly stronger notion of security known as IND-CCA or INDCCA2, “security under known ciphertext attack”. Nonces Nonces are random values enclosed with some formating data generated with a nonce generation algorithm N . They serve as a source of fresh, unguessable data exchanged in protocols. Nonce generation algorithms can be stateful, which somewhat complicates the appropriate de?nition of their security. We de?ne the advantage of an arbitrary algorithm A of breaking the security of nonce generation algorithm N as a probability of succeeding in the following game: k + l + 1 nonces are sequentially generated with N and then the algorithm A is run on the ?rst k and the last l nonces with the task to guess the value of k + 1-th nonce: Advnonce (A) = Pr[m, n, p ←? N (η) : A(m, p) = n] N If this advantage is negligible in η for every ppt algorithm A, then N is secure. In practice, this type of security is achieved by simply enclosing η long uniformly sampled string with formating data. 2.3 Confusion Freeness and Weak Key Authenticity

$

Neither the syntax of an encryption scheme nor the typical notions of security, such as the one de?ned above, say much about what happens if we attempt to decrypt an encryption with a key distinct from the decryption key. Syntax of an encryption scheme allows for such decryption to fail, but it does not insist on it. If it does not fail, notions of security forbid that the result is in any meaningful way related to the underlying plaintext — a ppt algorithm has no way of distinguishing it from any other potential plaintext with non-negligible probability. As a reader might already suspect, a failure to detect such situations would a?ect the completeness of an abstract model of cryptography. It is implicitly assumed that an abstract agent recognizes undecryptable encryptions in most if not all abstract models; if a ppt agent in the computational model is strictly weaker, then the abstract model would be incomplete. We might require that decrypting an encryption with independently generated fresh key fails with all but negligible probability (as a function of security parameter η). This property was de?ned in [MW04a] and called confusion freeness. It is su?cient to prove the completeness of an abstract model. Similar and independent de?nition can also be found in [AJ01]. 7

However, confusion freeness is a quite strong requirement on an encryption scheme. It turned out not to be necessary: a strictly weaker notion called weak key authenticity was de?ned and shown to be both necessary and su?cient for proving completeness [HG03]. Weak key authenticity requires only that an attempt to decrypt an encryption with incorrect decryption key fails with non-negligible probability.

3

3.1

The Abstract Model

Messages as Experiments

The act of creating a cryptographic message, in view of the probabilistic character of cryptographic algorithms, is a probabilistic experiment. Say the message is Ek (P(n, 0)). Without any contextual assumptions on the key k and nonce n, meaning that they should be freshly generated, this implies the following cryptographic experiment: [K ←? K(η); k ←? I(K); n ←? N (η); m ←? P(n, 0); e ←? E(k, m) : e] While it is easy to formalize the above notation for experiments directly, we skip it here. It should su?ce to say that an experiment is a sequence of actions delimited with semicolon; if the experiment has an output, then it is separated from preceding actions by a colon. Left arrows are assignment operators, sometimes decorated with $ to emphasize the use of probabilistic algorithms on the right hand side. Expanding the shorthand for probabilistic algorithms, the above experiment would take the form of [c1 ←? Coins; K ←? K(η, c1 ); k ←? I(K); c2 ←? Coins; n ←? N (c2 ); m ←? P(n, 0); c3 ←? Coins; e ←? E(k, m, c3 ) : e] We shall in the sequel assume that all experiments are so expanded, that ←? appears only at the left of Coins. 3.2 Experiments as Terms

$ $ $ $ $ $ $

Here we develop a more systematic notation for representing cryptographic probabilistic experiments, with well-known and widely used terms of ?rst-order logic. In logic every function symbol comes equiped with its arity and, optionally, can be marked as relational. We in addition mark some function symbols as probabilistic and some as parameterized. Here we list all vocabularies that will be used throughout this paper. Vocabularies: – Υlog is the vocabulary of logical constants, containing nullary symbols true, false and undef, the usual boolean operators and the equality =. 8

– Υexp contains unary symbols key, inv, fst, snd and nonce, binary decrypt and pair, and ternary encrypt. Symbols key, nonce and encrypt are marked as probabilistic and symbols key and nonce are also marked as parameterized. – Υconst contains nullary symbols for some constants, at least for bits 0 and 1. – Υf un contains unary relation symbols PriKey, PubKey, Ciphertext, Pair, unary function len and a binary relation symbol sameKey. – Υ = Υlog ∪ Υexp ∪ Υconst ∪ Υf un . For experiment-representing terms the vocabulary Υexp ∪Υconst ∪{undef} will su?ce, together with some set of additional constants to denote some coins. De?nition 2. Let C be a set of constants. The set of experiment-representing terms, in short e-terms, of vocabulary Υconst ∪ Υexp ∪ {undef} over C, is de?ned inductively as: – nullary symbols in Υconst and undef are e-terms; – if n-ary symbol f ∈ Υexp is not marked as probabilistic and t1 , . . . , tn are e-terms, then f (t1 , . . . , tn ) is an e-term; and – if n-ary symbol f ∈ Υexp is marked as probabilistic, t1 , . . . , tn?1 are e-terms and c ∈ C, then f (t1 , . . . , tn?1 , c) is an e-term. Given an assignment of in?nite strings to constants in C and a concrete value of security parameter η, we can assign a concrete string to every e-term. De?nition 3. Let t be an experiment-representing term of vocabulary Υconst ∪ Υexp ∪ {undef} over C, Π = (K, E, D) an encryption scheme, Σ = (P, F, S) a pairing scheme, N a nonce generation algorithm, and σ an assignment of in?nite strings to constants in C. Then a string t Π,Σ,N is de?ned inductively as follows η,σ (when Π, Σ, N are known, we drop them from the notation): – undef is interpreted as the failure string undef

η,σ

=⊥

– if g is neither marked as probabilistic nor marked as parameterized, then g(t1 , . . . , tn )

η,σ

= G ( t1

η,σ , . . . ,

tn

η,σ )

– if g is marked as probabilistic but not as parameterized, then g(t1 , . . . , tn?1 , c)

η,σ

= G ( t1

η,σ , . . . ,

tn?1

η,σ , σ(c))

– if g is marked as both probabilistic and parameterized, then g(t1 , . . . , tn?1 , c)

η,σ

= G (η, t1

η,σ , . . . ,

tn?1

η,σ , σ(c))

for every (g, G) ∈ {(key, K), (encrypt, E), (decrypt, D), (pair, P), (fst, F), (snd, S), (nonce, N )} and every c ∈ C. 9

Thus taking any e-term t, sampling for σ from the uniform distribution we $ obtain a probability distribution Pr σ ←? U : t η,σ ; varying η we obtain an ensemble. The assumptions on the encryption scheme force that Pr c, c ←? U : D(K(η, c), E(I(K(η, c)), m, c )) = m = 1 must hold for any message string m, while the confusion-freeness assumption forces $ Pr c, c , c ←? U : D(K(η, c), E(I(K(η, c )), m, c )) = ⊥ to be overwhelming for every message string m. Similar equivalences are forced by assumptions on the pairing function and projections. We show that these equivalences carry over to formalization by e-terms, for instance that Pr σ ←? U : decrypt(key(c1 ), encrypt(inv(key(c1 )), t, c2 )) must be overwhelming for every e-term t. De?nition 4 (Equivalence of E-Terms). Let T be a set of e-terms of vocab. ulary Υconst ∪ Υexp ∪ {undef} over C. Then = is the smallest equivalence over T induced by the clauses – for all e-terms tk , te , tm decrypt(tk , te ) = . tm ?ck , ce . tk = key(ck ) ∧ te = encrypt(inv(tk ), tm , ce ) undef otherwise

$ η,σ $

= t

η,σ

– for all e-terms t, tf , ts . ts if t = pair(tf , ts ) undef otherwise . . Let [t]= be the standard notation for the class of = equivalent terms. The above de?nition justi?es the common representation of cryptographic messages with terms without decrypt, fst and snd symbols: fst(t) = snd(t) = Corollary 2. For every e-term t there is an e-term t0 in which decrypt, fst and . snd do not occur and t0 = t. Finally, we will show that the equivalence just introduced is justi?ed by its computational interpretation. Lemma 1. Let t1 , t2 be experiment-representing terms of vocabulary Υconst ∪ Υexp ∪ {undef} over C, Π = (K, E, D) a confusion-free encryption scheme, Σ = . (P, F, S) a pairing scheme, N a nonce generation algorithm. If t1 = t2 , then Pr σ ←? U : t1 is overwhelming in η. 10

$ η,σ

. tf if t = pair(tf , ts ) undef otherwise

= t2

η,σ

. Proof. The proof is by induction on the de?nition of equivalence =. Use the assumptions on Π, Σ, the confusion freeness property of Π and the fact that negligible functions are closed under addition. Remark 3. A corresponding statement can be made in the case of weak key authenticity. Under this assumption, the statement of the above lemma becomes Pr σ ←? U : t1

$ η,σ

= t2

η,σ

is not negligible in η. The proof is (almost) the same. We shall often assume the following properties of encryption schemes: Pr c1 , c2 ←? U : K(c1 ) = K(c2 ) Pr ck , c1 , c2 ←? U : E(K(I(ck )), m, c1 ) = E(K(I(ck )), m, c2 ) are both negligible in η. We shall name these properties, which easily follow from the usual security notions such as IND-CPA, but are themselves much weaker, as “random keys” and “random encryption” properties. Lemma 2. Let Π be a confusion-free encryption scheme with random keys and random encryption properties, Σ a pairing scheme, N a secure nonce generation . . algorithm and t1 , t2 e-terms. If [t1 ]= = [t2 ]= then Pr σ ←? U : is negligible in η. Proof. By Lemma 1 and Corollary 2, it su?ces to show that Pr σ ←? U :

$ $ $ $

t1

Π,Σ η,σ

= t2

Π,Σ η,σ

t1 = t 2

is negligible for terms t1 and t2 in which decrypt, fst and snd do not occur. The rest of the proof is straightforward simultaneous induction on the structure of construction terms t1 and t2 . 3.3 Experiments as Structures

Given a set of abstract representatives of coins to interprete constants from C, . we can organize the e-terms modulo = to a ?rst-order structure. What it buys us is possibility to harness the well-developed theory of interactive algorithms of [BG04a,BG04b], which operate over such structures as their states. If elements of the structure are essentially equivalence classes of e-terms, . and = is closed under substitution, the interpretation of any function g in the vocabulary Υconst ∪ Υexp ∪ {undef} is naturally de?ned as . . . g([t1 ]= , . . . , [tn ]= ) = [g(t1 , . . . , tn )]= 11

The logical part of the structure is de?ned in the usual way. Some additional relations are added to the interpretation, re?ecting the assumptions on the tagging scheme T , holding in codomains of functions key, encrypt, pair and nonce. We proceed with a verbose de?nition of an isomorphism-closed classes of structures K. De?nition 5. Let K be an isomorphism closed class of Υ –structures such that X ∈ K if and only if there is a uniquely de?ned set CoinsX such that: – true, false and undef denote distinct elements; elements in domains and codomains of all logical constants except equality are logical elements in X; the interpretations of logical connectives in Υlog are the usual ones; – each k ∈ Υconst denotes a unique non-logical element, we denote the set of such elements Const; – domains and codomains of functions in Υconst and Υexp , and the set CoinsX contain non-logical elements only; – the non-logical part of X is freely generated with functions key, inv, encrypt, pair and nonce from CoinsX ∪ Const; – PriKey, PubKey, Ciphertext, Pair and Nonce hold on codomains of functions key, inv, encrypt, pair and nonce respectively, – sets CoinsX , Const, PriKey, PubKey, Ciphertext, Pair, Nonce are pairwise disjoint, and we de?ne Msg = Const∪PriKey∪PubKey∪Ciphertext∪Pair∪Nonce; – functions key, inv, encrypt, pair, nonce are injective, with the domains CoinsX , PriKey, PubKey × Msg × CoinsX , Msg × Msg, CoinsX , respectively; – decrypt, fst and snd are de?ned as fst(pair(m1 , m2 )) = m1 snd(pair(m1 , m2 )) = m2 decrypt(key(c1 ), encrypt(key(inv(c1 )), m, c2 )) = m for every m, m1 , m2 ∈ Msg and every c1 , c2 ∈ CoinsX ; elsewhere these functions take the value undef; – function len assigns an integer to each m ∈ Msg such that, assuming len(m1 ) = len(m1 ), len(m2 ) = len(m2 ), we have: len(pair(m1 , m2 )) = len(pair(m1 , m2 )) len(encrypt(inv(key(c1 )), m1 , c2 )) = len(encrypt(inv(key(c1 )), m1 , c2 )) len(m1 ) + len(m2 ) ≤ len(pair(m1 , m2 )) len(m1 ) ≤ len(encrypt(inv(key(c1 )), m1 , c2 )); for every m1 , m1 , m2 , m2 ∈ Msg and every c1 , c1 , c2 , c2 ∈ CoinsX ;if the argument is not in Msg, len takes the value undef. – relation sameKey holds in e1 , e2 i? e1 = encrypt(inv(key(c)), m1 , c1 ) and e2 = encrypt(inv(key(c)), m2 , c2 ) for some c, c1 , c2 ∈ CoinsX , m1 , m2 ∈ Msg. De?ning the structure, we have used e-terms with set of constants C = CoinsX . 12

What exactly is the relation of e-terms and structures just de?ned? Elements of CoinsX are not accessible by ground terms in a structure X ∈ K, and therefore e-terms cannot be directly evaluated in X. But if we expand the structure X with constant symbols denoting CoinsX , then non-logical elements can be seen as . classes of = equivalent terms. For X ∈ K, we will denote with X + its unique expansion with constants CoinsX denoting themselves in X + . Since the non-logical part of X is freely generated by key, inv, encrypt, pair, nonce from Const ∪ CoinsX , there is a unique ground term tX of vocabulary {key, inv, encrypt, pair, nonce} ∪ x X Υconst ∪ CoinsX denoting every non-logical x in X + . Denote with Tx the set of + X X . all ground terms denoting x in X . Then Tx is exactly [tx ]= . This reading of the de?nition allows us to attach the computational interpretation to elements of structures as well. De?nition 6. Let X ∈ K, x a non-logical element in X and σ an assignment of in?nite strings to CoinsX . Then x

X,η,σ

= tX x

η,σ .

If any of the parameters is determined by the context, we might suppress it and ultimately write t and x if all parameters are understood from the context. By Lemma 1 and Lemma 2, both equality and inequality on non-logical part are preserved with overwhelming probability. If we ?x some distinct coding of the logical elements, then we can extend the computational interpretation to all elements of the structure. The abstract interpretation will be preserved with overwhelming probability by the computational representation. Corollary 3. Let Π = (K, I, E, D) be a confusion-free encryption scheme with random keys and random encryption properties, Σ = (P, F, S) a pairing scheme, N a secure nonce generation algorithm, X ∈ K and t1 , t2 terms of X + . Then V al(t1 , X + ) = V al(t2 , X + ) if and only if t1 η,σ = t2 η,σ with overwhelming probability. 3.4 Experiments and Algorithms

If we wanted to capture full static logic of asymptotic computational cryptography, we would need much more involved logical constructions. But full static logic is not what we are after, capturing equality and inequality su?ces for our purposes. Equality and inequality, which means similarity, su?ces to determine the behavior of abstract interactive algorithms of [BG04a,BG04b]. Under the computational interpretation, concrete ppt Turing machines operating on concrete cryptographic messages can simulate abstract algorithms operating over structures representing such messages. A concrete ppt Turing Machine, run on a tape containing a ?nite set of cryptographic messages, can analyze the messages by running deterministic algorithms such as decryption D and projections of pairs F and S, testing parts of analyzed messages for equality etc. It can also create new messages by running probabilistic key and nonce generation algorithms K and N , encryption algorithm E, or deterministic algorithms such as the pairing algorithm P. 13

The fact that concrete ppt Turing machines can do essentially no more than the abstract algorithms will be forced by security assumptions on the encryption schemes. Abstract algorithms represent all possible internal actions of an algorithm with evaluation of terms, and external actions, such as receiving of input messages, with an answer function attached to a state. Internal memory of the abstract algorithm will be modeled with additional functions expanding the structures. The modeling choices we just made are quite obvious and su?cient for everything but coin ?ipping, e.g. creation of fresh nonces, encryptions etc. The behavioral theory of algorithms has a well developed theory of importing of fresh objects. Almost every non-trivial application of the theory use importing over a background structure. There is nothing fundamentally di?erent in extending the working space of an algorithm with a fresh atom used to build hereditarily ?nite sets, or with a fresh atom representing a fresh coin ?ip used for probabilistic functions. Only atoms that are not used in any meaningful way in the state can be imported, and the exact choice of the atom imported is irrelevant since they all produce isomorphic states. The isomorphism-closed class of structures K is a background classes with Atoms(X) = CoinsX for every X ∈ K. We will denote it by BCcpa . Let A be an ordinary interactive small–step algorithm with background BCcpa . In every state X, A evaluates a ?nite set of terms, possibly using results of interaction with its environment α, and ?nally, based of the result of the evaluation, generates an update set ?+ (X, α). We will make no limitations on A coins that can be imported by α, except the usual one that an imported coin must be a reserve atom.

De?nition 7. Let A be an ordinary interactive small-step algorithm with background BCcpa and X its state. Let Π an encryption scheme, Σ a pairing scheme and N a nonce generation algorithm. Then

– X

η,σ

is a concatenation of strings x

η,σ

for all accessible x ∈ X.

– A η is a Turing machine that evaluates computational interpretations of abstract terms evaluated by A.

Example 3. An abstract algorithm modeling the ?rst action of responder B in the Needham–Schroeder protocol: {kA , nA }k A ?? ? ? ? ? ? B ? ? ? ? ?B → ? {nA , nB }kA ←? ? ? ? ? ? ? ? ? ? ? ?? 14

is given with the following ASM program: let p = decrypt(B, in), kA = fst(p), nA = snd(p) in if PubKey(kA ) and Nonce(nA ) then import c1 , c2 in let nB = nonce(c1 ) in a := kA n := nA m := nB out(encrypt(kA , pair(nA , nB ), c2 )) The program is executed in a state X with background BCcpa and a context α. State X contains a constant B denoting the private key KB , and unde?ned constants n, m, and a. Context α is α = {(c1 , c1 ), (c2 , c2 ), (in, e1 ), (out[e2 ], ack)} for some c1 , c2 ∈ Reserve(X) and encryptions e1 , e2 ∈ X. Queries in and out are used for communicating with the environment, while c1 and c2 represent internal coin ?ips made by the algorithm—environment replies to queries c1 , c2 . The corresponding Turing machine A η operates on three tapes. The ?rst tape represents the internal memory of the algorithm and contains the string KB η,σ . The second one represents internal randomness needed by the algorithm, it is an in?nite sequence of random bits. The third tape represents interaction with an environment, containing e1 η,σ at beginning and e2 η,σ at the end of the calculation. The interpretation of actions of an abstract algorithm with experiments deserve some additional attention. The (abstract) work performed by an algorithm is measured in ground terms it evaluates. Evaluation of a term is inductively de?ned as: 1. interpretation of a background function, 2. interpretation of a foreground function, and 3. querying and receiving an answer from the environment. Work performed in (1) amounts to evaluation of the appropriate function G represented by a background function g. Foreground functions represent the internal memory of an algorithm, and therefore (2) is usually a simple memory lookup. Work performed by the environment (3) is not done by the algorithm, as it simply creates queries and uses the answers provided. This view has the following simple consequence when A is instantiated with a concrete implementation working on strings: Corollary 4. Let A be a small–step algorithm with background BCcpa , Π an encryption scheme, and Σ an pairing scheme. Then A η is a ppt Turing machine. 15

As already said, we will assume that every algorithm is capable of performing experiments on its own, and therefore we will pose no restriction on importing of fresh coins from the reserve of a state. However, algorithms might also receive, within a step, results of experiments from its environment, such as answers obtained by oracles. Speci?c modeling circumstances, such as a speci?c notion of security, will determine our restrictions on such answer functions. Theorem 4 (Completeness). Let Π = (K, I, E, D) be a confusion-free encryption scheme with random keys and random encryptions, Σ = (P, F, S) a pairing scheme and N a secure nonce generation algorithm. Furthermore, assume that for all but a ?nite number of values of η: – len(x) = len(y) in X i? | x η,σ | = | y η,σ | for every state X with background BCcpa ; – encryption scheme Π is equipped with a ppt algorithm that can distinguish two encryptions created with di?erent keys. Let X and Y be states with background BCcpa . If X ? Y then there is a ppt algorithm distinguishing X η and Y η with overwhelming probability. Proof. If X ? Y , then there are ground terms t1 and t2 such that V al(t1 , X) = V al(t2 , X) and V al(t1 , Y ) = V al(t2 , Y ). By Corollary 3, the equality and inequality of terms is preserved with overwhelming probability by the computational interpretation. Let A be an algorithm outputting true when t1 = t2 and false otherwise. Then A η distinguishes X η and Y η with overwhelming probability. By Corollary 4, A η is a ppt algorithm. 3.5 Abstract Notions of Security

Abstract interactive algorithms of [BG04a,BG04b] allow us to model the oracle adversary games de?ning cryptographic security notions directly on the abstract level! The security of an encryption scheme is completely characterized with the corresponding oracles attached to a ppt algorithm trying to break the security of an encryption scheme, as described in section 2.2. The oracles attached to the algorithm in a state with background BCcpa also perform experiments, and the results of the experiments are already representable by elements of the base set of the state. From the algorithm’s point of view, actions of oracles are actions of its environment. Interaction of an algorithm and environment is well studied in the behavioral theory of algorithm. It is represented with a collection of answer functions attached to a state completely characterizing all possible reactions of environment. We will show how to view de?nitions of the notions of security from section 2.2 abstractly as a set of abstract answer functions attached to a state of an algorithm. Recall the IND-CPA notion of security given by De?nition 2. Its abstract representation is as follows, minding that n, pk, eo(x, y) are queries asking the environment for fresh (internal) coin ?ips, a public key associated with the oracle or an encryption of one of the two messages respectively: 16

De?nition 8. Let X be a state with background BCcpa . Then – a context α of an ordinary interactive small–step algorithm A in X is INDCPA good if there are distinct elements ck , c1 , . . . , cn , ce , . . . , ce in the reserve 1 k of X such that α(ni ) = ci α(pk) = invX (keyX (ck )) encryptX (invX (keyX (ck )), xj , ce ) j α(eoj (xj , yj )) = undef

if lenX (xj ) = lenX (yj ) otherwise

for some xj , yj ∈ X, i = 1, . . . , n, j = 1, . . . , k; and – a context β of an ordinary interactive small–step algorithm A in X is INDCPA fake if there are distinct elements ck , c1 , . . . , cn , ce , . . . , ce in the reserve 1 k of X such that α(ni ) = ci α(pk) = invX (keyX (ck )) encryptX (invX (keyX (ck )), yj , ce ) j α(eoj (xj , yj )) = undef for some xj , yj ∈ X, i = 1, . . . , n, j = 1, . . . , k. Let A be the set of all IND-CPA good contexts and B the set of all IND-CPA fake contexts in state X. Then A, B is the abstract model of IND-CPA oracle interaction in X. Instantiations of these answer functions with concrete encryption schemes are exactly the experiments de?ned with the IND-CPA notion of security in De?nition 2. Both IND-CCA good and fake contexts are extension of IND-CPA good and fake contexts with additional queries and answers representing the decryption oracles α(do(xl )) = β(do(xl )) = decryptX (keyX (ck ), xl ) for some xl such that xl is not one of the answers to encrypt queries in α or β, for l = 1, . . . , m. The abstract model of interaction of IND-CPA notion of security induces an equivalence relation on states in the following way. We say that a small–step algorithm A reduces state X to state Y for answer functions α, β if X = A(0X , α) and Y = A(0Y , β). State X is reducible to Y for α, β if such a small–step algorithm exists. De?nition 9. Let A, B be the abstract model of IND-CPA interaction. Then cpa X is reducible to Y for IND-CPA, denoted with X ?→ Y , if X is reducible to Y for some α ∈ A and β ∈ B. If both α and β are parameterized with the same K oracle decryption key K, then we also write X ?→ Y . The equivalence induced by the reducibility relation for IND-CPA relation, cpa its transitive and symmetric closure, is denoted with X = Y . 17

if lenX (xj ) = lenX (yj ) otherwise

Lemma 3. Let Π be a confusion-free IND-CPA secure encryption scheme, Σ a pairing scheme and N a secure nonce generation algorithm. Let X and Y states cpa with background BCcpa . If X = Y , then X η is indistinguishable from Y η by probabilistic polynomial time algorithms with all but negligible probability. Proof. Since computational indistinguishability is an equivalence, it is su?cient cpa to show that X η ≈ Y η when X ?→ Y . We argue by contradiction. Suppose there is a ppt algorithm A distinguishing X η and Y η with non-negligible probability. We will use this algorithm to distinguish oracles characterizing IND-CPA security. Denote with α an IND-CPA good context and with β an IND-CPA fake context such that for some algorithm A we have X = A(0X , α) and Y = A(0Y , β). Run the algorithm A η with an IND-CPA oracle to create X η,σ or Y η,σ , depending on whether you are provided with a good oracle or a fake one. Run A on the resulting state. If A can distinguish X η,σ from Y η,σ with nonnegligible probability, then we can break IND-CPA security of the encryption scheme used. The above lemma tells us that certain challenges are indistinguishable as a simple consequence of the notion of security. If two inputs of a challenge can be generated by the same abstract algorithm, but using two di?erent oracles, then it is clear that this algorithm, if successful, would break security of the underlying encryption scheme. Remark 4. Given some enumeration of coins c1 , c2 , . . . in a state X, we will often use the following notation for elements of X: ni for nonceX (ci ), Ki for keyX (ci ), ki for invX (keyX (ci )), m1 , m2 for pairX (m1 , m2 ) and {m}i j for K encryptX (kj , m, ci ). If a state has a single nullary foreground symbol, then we will identify the state with the unique element the symbol is denoting in it. E.g. a state X with f denoting encryptX (invX (keyX (c2 )), keyX (c1 ), c3 ) for some coins c1 , c2 , c3 ∈ Reserve(X) is identi?ed with {K1 }3 2 . K Example 4. We will show that K2 K1 {{K1 }4 2 , K3 }5 1 ?→ {{0}4 2 , K3 }5 1 ?→ {0}5 1 K K K K K for IND-CPA security (assuming that 0 denotes a zero string of an appropriate length). Let A1 and A2 be algorithms with programs Π1 and Π2 : Π1 = import c1 , c3 , c5 let K1 = key(c1 ), K3 = key(c3 ), k1 = inv(K1 ) in f := encrypt(K1 , pair(eo(K1 , 0), K3 ), c5 ) import c2 , c3 , c4 let K2 = key(c2 ), K3 = key(c3 ), k2 = inv(K2 ) in f := eo(pair(encrypt(k2 , 0, c4 ), K3 ), 0) 18

Π2

=

and let IND-CPA positive α1 , α2 and IND-CPA negative β1 , β2 be α1 = {(c1 , c1 ), (c3 , c3 ), (c5 , c5 ), (eo[K1 , 0], {K1 }4 2 )} K β1 = {(c1 , c1 ), (c3 , c3 ), (c5 , c5 ), (eo[K1 , 0], {0}4 2 )} K α2 = {(c2 , c2 ), (c3 , c3 ), (c4 , c4 ), (eo[ {0}4 2 , K3 , 0], { {0}4 2 , K3 }5 1 )} K K K β2 = {(c2 , c2 ), (c3 , c3 ), (c4 , c4 ), (eo[ {0}4 2 , K3 , 0], {0}5 1 )} K K Then A1 (0, α1 ) A2 (0, α2 ) = = {{K1 }4 2 , K3 }5 1 K K {{0}4 2 , K3 }5 1 K K A1 (0, β1 ) A2 (0, β2 ) = = {{0}4 2 , K3 }5 1 K K {0}5 1 K

By Lemma 3, we can conclude that Adv(A) = Pr K1 , K2 , K3 ←? K(η); e ←?: Ek1 (P(Ek2 (K1 ), K3 )) : A(e) = 1 ? Pr K1 ←? K(η); e ←? Ek1 (0 ) : A(e) = 1 is negligible for every ppt algorithm A (we use ki = I(Ki )).

$ $ $ $

4

4.1

Soundness and Completeness of the Abstract Model

Indistinguishability

If we prove that state X and Y are indistinguishable by small–step algorithms, what have we proved? We hope that than there is no probabilistic polytime algorithm that can distinguish strings produced by experiments encoded by X and Y with all but negligible probability. For the IND-CPA notion of security, the background class BCcpa , an encryption scheme Π, a pairing scheme Σ and a nonce generation algorithm N , we have three equivalence relations on abstract states with background BCcpa representing experiments: 1. computational indistinguishability; 2. abstract indistinguishability; and 3. abstract reducibility. If X η and Y η are indistinguishable by ppt algorithms, we write X ≈ Y . Computational indistinguishability is the semantical relation on states, de?ned independently from our formalism in terms of capabilities of probabilistic polynomial time Turing machines. If X and Y are indistinguishable by small–step algorithms, we write X ? Y . Abstract indistinguishability articulates our intention about what an encryption scheme should achieve. It can also be seen as the power explicitly given to an agent by an encryption scheme: if an agent can distinguish two states with an abstract algorithm, then she can use an instantiation of the program to distinguish 19

Π

instantiations of the states, all with the concrete instantiated encryption scheme. This property is usually called completeness, and it can be phrased as “whatever an abstract algorithm can do, a concrete instantiation can do with overwhelming probability as well”. The proof given in Theorem 4 is quite straightforward, but it involves some simple reasoning about probabilities. This is necessary, since it relates an abstract relation with semantics de?ned in terms of ppt algorithms. We get

Π

X ? Y ? X ≈ Y, or equivalently X ≈ Y ? X ? Y.

Π

(1)

Abstract reducibility tells us what a concrete ppt algorithm cannot do as a direct consequence of the notion of security. The proof given in Lemma 3 is again very simple, it is nothing more than expressing what is the true meaning of a particular notion of security. We get X = Y ? X≈Y From equations (1) and (2), we have X = Y ? X≈Y ? X?Y If we could relate abstract notions of equivalence by showing that X ? Y ? X = Y, we could, using (3), conclude that all three notions are equivalent X ? Y ? X = Y ? X ≈ Y. The theorem establishing (4) is the essence of the computational soundness of abstract wrt computational cryptography. It is also the most di?cult one to prove. However, it is expressed and proved completely in abstract terms, with no mention of ppt Turing machines and their probabilities to distinguish concrete strings. The set of submessages of a message x in state X, denoted with SubX (x), is de?ned inductively as m1 ∈ SubX (m1 ), m2 ∈ SubX (m2 ) =? m1 , m2 ∈ SubX (pairX (m1 , m2 )) ms ∈ SubX (ms ) =? SubX (encryptX (invX (keyX (ck )), ts , ce )) for all messages m1 , m1 , m2 , m2 , ms , ms ∈ MsgX and all coins ce , ck ∈ CoinsX . Key k is the decryption key of encryption e and message m is the subject in a state X if e = encryptX (k, m, c) for some coin c. The set of used messages in a state X are the smallest set of messages closed under submessages and decryption keys that include all exposed messages in X. 20

cpa Π cpa cpa Π cpa Π

(2)

(3)

(4)

A state X is said to be in (IND-CPA) normal form if every accessible encryption with inaccessible decryption key has zero string (of appropriate length) as a subject. Let A, B be the model of interaction for IND-CPA security in 0X . If an inaccessible key k is not a subject of any used message in a state X, then X can be constructed from 0X by a small–step algorithm A and some α ∈ A. If A is run on 0X , β for some β ∈ B, we can produce a state X in which k encrypts only zeros, like we did in Example 4. Lemma 4. Let X be a state with background BCcpa and K a decryption key cpa not occurring as a submessage of any used encryption in X. Then X ?→ X for some state X with the same background reduct and the same accessibility of nonces and keys such that – all keys and nonces are accessible with the same terms in both X and X ; – key K encrypts only zero strings in X ; and – if encryption key k1 encrypts decryption key K2 or nonce n in X , then k1 encrypts K2 or n in X as well. Now we have everything we need to prove that acyclic states are reducible to normal form. Lemma 5. Let X be an acyclic state with background BCcpa and accessible all exposed elements. Then X is IND-CPA reducible to its normal form. Proof. Enumerate inaccessible decryption keys such that encryption key kj does not encrypt Ki if i ≤ j. Since the state is acyclic, such numeration is possible. We will reduce X in n steps to a state Xn such that 0 is the only used subject encrypted by an inaccessible encryption key. The proof is by induction on the enumeration of inaccessible decryption keys. Key K1 is inaccessible and does not occur as a submessage of subject of any used encryption in X. By Lemma 4, then cpa there is a state X1 such that X0 ?→ X1 , k1 does not encrypt any decryption key in X1 and encrypts in X1 is a subrelation of encrypts in X. Hence, the enumeration of keys in X is good for X1 . Since key k1 does not encrypts any decryption key in X1 , there is no key in X1 that encrypts K2 in X1 . But then K2 satis?es the condition of Lemma 4 in X1 and we can make another step of the induction. In Xn , subject of every undecryptable encryption is zero. Thus Xn is in the normal form. An example of a reduction of a (cyclic) state to its normal form is given in Example 4. Since a normal form is a representative of its similarity class, we have: Corollary 5. Let X and Y be acyclic states with background BCcpa and accescpa sible all exposed elements. If X ? Y then X = Y . 21

Theorem 5 (Soundness). Let Π be an IND-CPA secure encryption scheme, Σ a pairing scheme, N a nonce generation algorithm, and X and Y acyclic states with background BCcpa . If X ? Y , then X η and Y η are indistinguishable by probabilistic polynomial time algorithms. Proof. We will assume that all exposed elements are accessible in both states. If a state contains exposed but inaccessible elements, replace it with a state obtained by unde?ning all foreground functions on such elements. The resulting state is clearly computationally indistinguishable from the original one, it provides the cpa same information to the intruder. By Corollary 5, we have X = Y . Finally, by Lemma 3, X η and Y η are indistinguishable. 4.2 Accessibility

If an element x is not accessible by a term in X, α for some state X with background BCcpa and α whose codomain is in the reserve of the state, can we conclude that no ppt algorithm can output x η when run on X η with nonnegligible probability? The similar theorem was proved for an Abadi–Rogaway language in [MW04b]. We will extend the soundness result of the previous subsection to accessibility here. Lemma 6. Let X be a state with background BCcpa and x ∈ X with non-empty support in X: SupX ({x}) = ?. Then Pr σ ←? U : A = x is negligible for every ppt algorithm A. Intuitively, this means that a ppt algorithm cannot guess an independently created key, nonce or encryption if no data is provided to it. Lemma 7. Let Π be an IND-CCA secure encryption scheme and X, X nonK isomorphic states with background BCcpa such that X ?→ X for IND-CCA contexts α and β. If some ppt algorithm A can produce x X,η,σ with nonnegligible probability when run on X η,σ , then – decryption key K is not a submessage of x in X (key invX (K) can occur as a submessage or as an encryption key); – there is a term t such that x = V al(t, 0X , α) and A can produce x X ,η,σ with non-negligible probability when run on X η,σ for x = V al(t, 0X , β); – x is accessible in X i? x is accessible in X . Proof. We argue by contradiction. Suppose that K is a submessage of x in X. All decryption keys are accessible in 0X , α except K, and every encryption created by encryption oracle in α does not contain K in the subject. Hence we can use IND-CCA decryption oracle and decryption with accessible keys to retrieve K X,η,σ from x X,η,σ . But then we can distinguish X η,σ and X η,σ with non-negligible probability, which is a contradiction by Lemma 3. 22

$ X,η,σ

Since K can only occur in x as an encryption key, there is a term t such that x = V al(t, α). Suppose that A can produce x X ,η,σ with negligible probability only. Then Pr σ ←? U : t

$ η,σ

= A( X

η,σ )

? Pr σ ←? U : t

$

η,σ

= A( X

η,σ )

is non-negligible and can be used to distinguish states X and X , which is a contradiction. The last part is a simple consequence of X and X being indistinguishable. Lemma 8. Let Π be an IND-CCA secure encryption scheme and X a state with background BCcpa in the IND-CPA normal form. If x is not accessible in X, then x X,η,σ is not accessible with non-negligible probability to ppt algorithms. Proof. If x is not accessible in X, then some inaccessible key K must occur as a submessage in x. Use IND-CCA decryption oracle and keys not used in X to retrieve it and break the IND-CCA security. Theorem 6. Let Π be an IND-CCA secure encryption scheme and X a state reducible to its normal form. If x ∈ X is not accessible in X to a small-step importing algorithm, then Pr σ ←? U : A( X is negligible for every ppt algorithm A. Proof. Let Xn be the normal form of X and K1 K2 K3 Kn X ?→ X1 ?→ X2 ?→ . . . ?→ Xn for some inaccessible decryption keys K1 , . . . , Kn . Let x, x1 , x2 , . . . , xn be elements from Lemma 7. Then xn is inaccessible in Xn . By Lemma 8, xn Xn ,η,σ is not accessible by ppt algorithms with non-negligible probability, which is a contradiction by Lemma 7.

$ η,σ )

= x

X,η,σ

5

Model–based Testing of Protocols

In this section we will show how to encode ASM programs working over BCcpa in Spec# and how to use SpecExplorer to explore all possible execution traces for a bounded number of roles and agents. The interactive algorithms of [BG04a,BG04b] are implemented in AsmL and Spec#. SpecExplorer is a tool developed at Microsoft Research for exhaustive exploration of (?nitized) state spaces of speci?cations written in AsmL or Spec#, in order to test an implementation for conformance and to generate unit tests. We have found that it can be e?ectively used to explore state spaces of protocol adversary situations, given a ?nite number of roles ensuring that the state space is ?nite. Our analysis shows that this exploration also has direct computational signi?cance. 23

class Coins {} structure Message{ public virtual int len(){ return 1; } case Nonce{ private Coins c; } case Pair{ public Message fst; public Message snd; public override int len(){ return fst.len() + snd.len(); } } case PrivateKey{ private Coins c; public PublicKey inv(){ return PublicKey(this); } } case PublicKey{ private PrivateKey sk; } case Encryption{ public PublicKey pk; private Message subject; private Coins c; public Message decrypt(PrivateKey sk) require sk.inv() == pk; { return subject; } public bool sameKey(Encryption e){ return pk == e.pk; } public override int len(){ return subject.len() + 1; } } public PrivateKey key(){ return PrivateKey(new Coins()); } public Encryption encrypt(PublicKey pk, Message subject) { return Encryption(pk, subject, new Coins()); } public Pair pair(Message f, Message s){ return Pair(f, s); } }

Fig. 1. BCcpa background encoding in Spec#

24

5.1

Encoding in Spec# and SpecExplorer

The original idea of modeling abstract properties of cryptographic primitives by an object–oriented programming language is from [RRS03]. An encoding of the BCcpa background in Spec# is given in Figure 1. The encoding should be clear to a reader with some basic understanding of accessibility modi?ers private and public in OO programming languages. A more elaborate discussion of the similar encoding in AsmL language can be found in [RRS03]. An honest role will be represented with a Spec# encoding of an ASM program operating over BCcpa background, such as the one given in example 3. On the other hand, the intruder will not be represented explicitly in the model with a concrete program. We will use the exploration capabilities of SpecExplorer to explore all possible execution paths with a given set of honest roles. It is not very di?cult to teach SpecExplorer to completely analyze a message created by an honest role, but we have a dramatically di?erent situation when it comes to creating a message that would be accepted by a role, forcing it to make a step and possibly output a fresh message. The set of states of an honest role is closed under isomorphisms and therefore in?nite. We will look into a very common class of protocols in which the set of messages that can be created by an intruder and accepted by an honest role is in?nite, but representable with a ?nite set of messages. Every step that a role can make will produce a state isomorphic to one of the states obtained by running the role with one of the representative messages. A role of a protocol exposes a very simple interface to the outside world. It analyzes an input message and, if certain conditions are satis?ed, outputs a message. If conditions are not met, it typically hangs, not producing a new messages regardless of any future inputs. If A is an action with a message m given as input in a state X and an answer function α with only reserve elements in its codomain, then τA (X, m, α) is a state obtained by ?ring A in X, m, α. An action of such role is called simple if it checks the type of all submessages of an input message. This means that any encryption must be decrypted, any pair must be analyzed, and type of any nonce and key must be checked. A protocol is simple if all actions of its roles are simple. Theorem 7. Let A be a simple action and I an intruder with a ?nite set of messages accessible by ground terms. Then there is a ?nite set of messages M accessible to I in X, α such that for every message m accessible to I in X, β there is m ∈ M such that τA (X, m, α) ? τA (X, m , β) = Proof. If action A is simple, then atomic support of m in X SupX (m ) is ?nite and bounded. It su?ce to import |SupX m | fresh coins from the reserve of X and create all messages of a ?xed submessage structure using freshly created coins and known messages. We use the above fact, together with theorems 1 and 6, to produce abstract representations of all non-negligible computational traces representable by the BCcpa background for a bounded number of honest roles. 25

Fix a simple protocol P . Let k be a maximum number of coins in support of any message accepted by an action of an honest role in P . At the initialization phase, SpecExplorer creates a ?xed number of honest agents and corrupted keys. In each subsequent step, SpecExplorer imports k fresh coins from the reserve, and creates all messages from fresh coins and already know messages that could potentially be accepted by some honest role. Each role is ran with each such message used as its input by SpecExplorer, thus producing a set of all reachable states in the model. The exact order in which states are explored is non-deterministic, although the tool allows di?erent priorities to be assigned to states. A role accepting an input message, can possibly output a message. The output message gets analyzed by the tool, thus updating the internal memory of the intruder. At the end of each step, the protocol guarantees are checked in new states by the tool. If any of protocol guarantees is not satis?ed, the exploration is aborted and a graph with explored states, including the one in which the guarantee is not ful?lled, is rendered by the tool. The trace that resulted with the bad state can be explored and studied using the tool. One optimization of the exploring process can be achieved by a grouping of states and further exploration of a single state representative of a group. The exploration space can be dramatically reduced for an appropriate grouping, but an optimal grouping is not always easy to ?nd. In our case, isomorphism seems like a good choice of grouping relation on states. Since SpecExplorer does not have a built-in option of grouping of isomorphic states, we use an ad–hoc coding of states resulting in a grouping relation ?ner than isomorphism, but still signi?cantly reducing the exploration space. Example 5 (Lowe’s attack on Needham–Schroeder protocol). One example of a simple protocol in the above sense is the public–key variant of the Needham– Schroeder authentication protocol. The ?aw found by Lowe is easily (re)discovered by SpecExplorer, usually in less than 100 explored states. The model is initially in a state marked as Initial in Figure 2. The ?rst step of initialization is performed by calling CreateAgentFactory, which creates an object capable of creating honest agents of the protocol. The next invocation of CreateAgents(2,1) creates two honest agents and one corrupted private key. The internal memory of the intruder is enriched with public keys of the honest agents and the private corrupted key. The resulting state is marked with Agents Created. The exploration process starts here. SpecExplorer now can create fresh roles of already created honest agents using CreateInitiatorRole and CreateresponderRole, or run already an created role using RunRole. The parameters for RunRole are picked up from the ?nite set of representative of messages using the Theorem 7. The Lowe’s attack on the protocol is found after creating one initiator and one responder role, and then calling RunRole four times with the appropriate parameters. In the resulting state, protocol guarantee is violated and the exploration process is terminated. The resulting state is clearly marked and the trace leading to the state is included in its description in the exploration graph in Figure 2. 26

Initial CreateAgentFactory()

Agent factory created

CreateAgents(2, 1) Agents created CreateInitiatorRole(.. CreateResponderRole(.. CreateInitiatorRole(.. Roles created CreateResponderRole(.. CreateInitiatorRole(.. CreateInitiatorRole(.. RunRole(nsrr1, ev15:.. Roles created Roles created RunRole(nsrr1, ev15:..

Fig. 2. SpecExplorer discovers Lowe’s attack on the Needham–Schroeder protocol

NSResp(A, B, m, n): {B, m}_A --> {m, n}_B NSResp(A, E, n, m): {E, n}_A --> {n, m}_E

27

RunRole(nsrr1, ev20:.. RunRole(nsrr1, ev20:.. NSResp(B, A, m, n): {A, m}_B --> {m, n}_A RunRole(nsir5, null) NSResp(B, E, n0, n): {E, n0}_B --> {n0, n}_E RunRole(nsir5, null) NSInit(A, E, m): - --> {A, m}_E RunRole(nsrr1, ev38:.. RunRole(nsrr1, ev38:.. NSResp(B, A, m, n): {A, m}_B --> {m, n}_A RunRole(nsir5, ev77:.. NSInit(A, E, m, n): {m, n}_A --> {n}_E RunRole(nsir5, ev28:.. RunRole(nsrr1, ev84:..

Roles created

Roles created

Roles created

ForceLoweAttack()

NSResp(B, E, n, m): {E, n}_B --> {n, m}_E

RunRole(nsir5, ev39:.. NSInit(A, E, m, m): {m, m}_A --> {m}_E RunRole(nsrr1, ev51:.. RunRole(nsrr1, ev51:.. NSResp(B, A, n, n0): {A, n}_B --> {n, n0}_A

NSInit(A, E, n0): - --> {A, n0}_E

NSInit(A, E, n0, n1): {n0, n1}_A --> {n1}_E

NSResp(B, E, n1, n0): {E, n1}_B --> {n1, n0}_E

Guarantee BROKEN! Role: NSResp(B, A, m, n) NSInit(A, E, m, n): - --> {A, m}_E NSResp(B, A, m, n): {A, m}_B --> {m, n}_A NSInit(A, E, m, n): {m, n}_A --> {n}_E NSResp(B, A, m, n): {n}_B --> -

References

[ABS05] Pedro Ad?o, Gergei Bana, and Andre Scedrov. Computational and infora mation theoretic soundness and completeness of formal encryption. In 18th IEEE Computer Security Foundations Workshop – CSFW 2005, 2005. [AJ01] Martin Abadi and Jan J¨rjens. Formal eavesdropping and its computational u interpretation. In Theoretical Aspects of Computer Software (4th International Symposium, TACS ’01), volume 2215 of LNCS, 2001. [AR02] Martin Abadi and Phillip Rogaway. Reconciling two views of cryptography (The computational soundness of formal encryption). Journal of Cryptology, 15(2):103–127, 2002. [Ban04] Gergei Bana. Soundness and Completeness of Formal Logics of Symmetric Encryption. PhD thesis, University of Pennsylvania, 2004. [BDPR98] M. Bellare, A. Desai, D. Pointcheval, and P. Rogaway. Relations among notions of security for public–key encryption schemes. In CRYPTO ’98, volume 1462 of LNCS, 1998. [BG00] Andreas Blass and Yuri Gurevich. Background, reserve, and Gandy machines. In Proceedings of CSL ’00, volume 1862 of LNCS, 2000. [BG03] Andreas Blass and Yuri Gurevich. Algortihms: A quest for absolute definitions. Bulletin of the European Association for Theoretical Computer Science, (81):195–225, October 2003. [BG04a] Andreas Blass and Yuri Gurevich. Ordinary interactive small–step algorithms I. Technical Report MSR-TR-2004-16, Microsoft Research, 2004. [BG04b] Andreas Blass and Yuri Gurevich. Ordinary interactive small–step algorithms II. Technical Report MSR-TR-2004-88, Microsoft Research, 2004. [Gur00] Yuri Gurevich. Sequential abstract state machines capture sequential algorithms. ACM Transactions on Computational Logic, 1(1):77–111, July 2000. [Gur05] Yuri Gurevich. Interactive algorithms 2005. Technical Report MSR-TR2005-73, Microsoft Research, 2005. [HG03] Omer Horvitz and Virgil Gligor. Weak key authenticity and the computational completeness of formal encryption. In Crypto 2003, volume 2729 of LNCS, 2003. [MW04a] Daniele Micciancio and Bogdan Warinschi. Completeness theorems for the Abadi-Rogaway language of encrypted expressions. Journal of Computer Security, 12(1):99–130, 2004. [MW04b] Daniele Micciancio and Bogdan Warinschi. Soundness of formal encryption in the presence of active adversaries. In Theory of cryptography conference - Proceedings of TCC 2004, volume 2951 of LNCS, 2004. [RR05] Dean Rosenzweig and Davor Runje. Some things algorithms cannot do. Technical Report MSR-TR-2005-52, Microsoft Research, 2005. [RRS03] Dean Rosenzweig, Davor Runje, and Neva Slani. Privacy, abstract encryption and protocols: an ASM model – Part I. In ASM 2003, volume 2589 of LNCS. Springer-Verlag, 2003. [AsmL] The AsmL webpage. http://research.microsoft.com/asml/. [Spec#] The Spec# webpage. http://research.microsoft.com/specsharp/. [SpecExp] The SpecExplorer webpage. http://research.microsoft.com/specexplorer/.

28

- a taxonomy of model-based testing
- Model-based testing of a highly programmable system
- Model-based Testing of Service Infrastructure Components
- Reference “Automated Model-Based Testing of
- Model Based Requirements Analysis and Testing of Automotive Systems
- Macro-model-based stress testing of Basel II capital requirements
- Model-based Testing of Reactive Systems A Seminar Volume 19 Model checking
- Games–Based Model Checking of Protocols counting doesn’t count
- Model-Based Testing of Automotive Electronics 1. Role of Testing throughout the Development
- Systematic Model-Based Testing of Embedded Automotive

更多相关文章：