Modeling Random Oracles Under Unpredictable Queries
 2 Citations
 909 Downloads
Abstract
In recent work, Bellare, Hoang, and Keelveedhi (CRYPTO 2013) introduced a new abstraction called Universal Computational Extractors (UCEs), and showed how they can replace random oracles (ROs) across a wide range of cryptosystems. We formulate a new framework, called Interactive Computational Extractors (ICEs), that extends UCEs by viewing them as models of ROs under unpredictable (aka. highentropy) queries. We overcome a number of limitations of UCEs in the new framework, and in particular prove the adaptive RKA and semiadaptive KDM securities of a highly efficient symmetric encryption scheme using ICEs under key offsets.
We show both negative and positive feasibility results for ICEs. On the negative side, we demonstrate ICE attacks on the \(\mathsf {HMAC}\) and \(\mathsf {NMAC}\) constructions. On the positive side we show that: (1) ROs are indeed ICE secure, thereby confirming the structural soundness of our definition and enabling a finer layered approach to protocol design in the RO model; and (2) a modified version of Liskov’s Zipper Hash is ICE secure with respect to an underlying fixedinputlength RO, for appropriately restricted classes of adversaries. This brings the first result closer to practice by moving away from variableinputlength ROs. Our security proofs employ techniques from indifferentiability in multistage settings.
Keywords
Random oracle Unpredictability UCE RKA security KDM security Zipper Hash Indifferentiability Multistage security1 Introduction
1.1 Background
Since their formal introduction by Bellare and Rogaway [BR93], random oracles (ROs) have found many applications across a wide range of cryptographic protocols. However, due to an uninstantiability result of Canetti et al. [CGH98], which shows that certain (artificial) protocols become insecure as soon as the random oracle is replaced by any concrete hash function, reliance on ROs has also become somewhat debatable.
Two lines of research have been directed at dealing with such uninstantiability results. One is to construct standardmodel counterparts of cryptographic primitives designed in the RO model (ROM). This approach comes with the drawback that the resulting cryptosystems often tend to be complex and achieve a lower level of security and/or efficiency. A second, more modular, approach aims to formulate abstractions of the proofcentric properties of random oracles such as extractability, programability, or nonmalleability [Can97, CD09, Nie02, CD08, BCFW09]. Assuming that a hash function meets the introduced model, one proceeds to show that it can safely replace the random oracle in a protocol. These formalizations, however, have only been successful to a limited extent, and the question of finding a flexible and general framework that could be applied across a broad range of security goals and protocols remained open until recently.
1.2 UCE Security
Bellare, Hoang, and Keelveedhi (BHK) [BHK13a] revisit the above questions and present a powerful framework called Universal Computational Extractors (UCEs) that allow to securely instantiate random oracles in an interesting and diverse set of applications. These include, among other things, security under keydependentmessage (KDM) attacks, security under relatedkey attacks (RKAs), simultaneous hardcore bits, point function obfuscators, garbling schemes, proofs of storage, deterministic encryption, and messagelocked encryption, thereby going far beyond what was previously possible.
 (1)

Provide the adversary with the hash key and ask it to distinguish an oracle implementing the hash function from one implementing the random oracle. This approach immediately fails as this game can be trivially won with the knowledge of the hash key by computing a hash value and checking the answer against the oracle’s answer for the same query.
 (2)

Adopt the above approach, but now hide the hash key. This leads to PRF security—for which feasibility results are known—but is not useful in the context of hashing as the hash key is typically publicly known.
BHK overcome the above shortcomings by splitting the attacker into two parts and constraining the communication between the two. The first UCE attacker does not get to see the hash key, but has oracle access to either the hash function under a random key or the random oracle according to a random bit. The second attacker, on the other hand, does get to see the hash key, but can no longer access the oracle, and it has to guess the bit; see Fig. 1 (left). The two stages of the adversary can communicate only in restricted ways since arbitrary communication would lead to an attack similar to that given above for formulation (1).
More formally, for a keyed hash function \(\mathsf {H}\), \(\mathrm {UCE}\) security is defined via a twostage game consisting of algorithms \(S \) and \(D \), called the source and the distinguisher respectively, as follows. In the first stage, the source is given access to an oracle Hash that depending on a random bit b implements either the random oracle or the concrete hash function \(\mathsf {H}\) under a random hash key \({{hk}} \). The source terminates by outputting some leakage L, which is then communicated to the secondstage distinguisher \(D \). In addition to leakage L, the distinguisher also gets the hash key \({{hk}} \) as input. The distinguisher’s task is to guess b, i.e., guess whether the source was talking to the random oracle or the hash function. The \(\mathrm {UCE}\) advantage of the pair \((S, D)\) is defined as usual to be the probability of correctly guessing the bit b scaled away from onehalf. We refer the reader to the original work [BHK13b] for an excellent overview of this approach to modeling hashfunction security.
To see that without further restrictions \(\mathrm {UCE}\) security cannot be achieved, consider a source that leaks one of its oracle queries together with the corresponding oracle answer to the distinguisher. The distinguisher then simply recomputes the hash value on the queried point—the distinguisher knows the hash key—and compares it to the leaked value.
In their original work, BHK [BHK13a] define two restrictions on sources: computational unpredictability and computational reset security. In the computational unpredictability game, it is required that when the source is run with a random oracle its leakage does not computationally reveal any of its queries. This is formalized by requiring that the probability of any efficient predictor \(P\) in guessing a query of \(S \) when given L is negligible.
The class of computationally unpredictable sources is denoted by \(\mathcal {S} ^\mathrm {cup}\), and the resulting \(\mathrm {UCE}\) security \(\mathrm {UCE}[\mathcal {S} ^\mathrm {cup}]\) (aka. \(\mathrm{UCE}1\)) of a hash function is defined by requiring the advantage of any efficient pair \((S,D)\) with an unpredictable \(S \in \mathcal {S} ^\mathrm {cup}\) in the \(\mathrm {UCE}\) game to be negligible. Reset security imposes a weaker restriction on the source class and leads to the stronger UCE2 notion.
UCE security has been the subject of many recent studies. Brzuska, Farshim, and Mittelbach (BFM) [BFM14] show that, under new cryptographic assumptions, these restrictions are insufficient for a feasible definition. More precisely, assuming the existence of indistinguishability obfuscators [BGI+01, GGH+13], BFM show that the \(\mathrm {UCE}[\mathcal {S} ^\mathrm {cup}]\) security of any hash function can be broken in polynomial time. To overcome this attack, BFM [BFM14] (and subsequently BHK in an updated version of their paper [BHK13b]) propose a statistical notion of unpredictability whereby the predictor can even run in unbounded time. Following the attack, BHK also refine the \(\mathrm {UCE}\) notions based on computational unpredictability and introduce the classes of bounded parallel and split sources.^{1} BFM show that security against bounded parallel source is also infeasible [BFM14], and recently attacks against split sources have also been shown [BST15].
On the positive side, Brzuska and Mittelbach [BM14b, BM15] show how to construct UCEs for the class of strongly unpredictable and statistically unpredictable sources for bounded number of queries. Bellare et al. [BHK14] develop domain extenders for UCEs, and Bellare and Hoang [BH15] construct deterministic PKEs from UCEs for statistically unpredictable sources and lossy trapdoor functions. BFM [BFM14] have shown that the existence of obfuscationbased attacks against statistically unpredictable sources violates wellknown impossibility results. A number of recent works have shown how to use UCEs as RO replacements in other protocols [MH14, BK15, DGG+15].
1.3 Interactive Computational Extractors
Given the development of UCEs, defining an extended model which meets the abovementioned specifications is an intricate task. Indeed, well before the emergence of obfuscationbased attacks, BHK [BHK13b, p. 9] warned that extending UCEs to an interactive setting is “a dangerous path to tread.” As an example, assume that we introduce a bidirectional communication channel between the distinguisher and the source so that our adaptivity targets are met. This extension can be shown to fall prey to somewhat nontrivial attacks that utilize generalpurpose multiparty computation (MPC) protocols. Suppose the source \(S \) holds a random input x whose hash is y, and \(D \) holds \({{hk}} \). The two parties then run an MPC protocol to compute the Boolean value \(y=\mathsf {H}({{hk}},x)\). The distinguisher finally returns this value as its guess. This attack would meet any reasonable notion of computational unpredictability since the security of the MPC protocol would ensure that the parties learn no more than what can be deduced from their individual private inputs.^{2} Allowing hash queries to depend on the hash key \({{hk}} \) is also challenging since similarly to approach (1) above access to both \({{hk}} \) and the hash oracle would trivialize the notion. For similar reasons, formulating a UCElike model for unkeyed hash functions is also nontrivial. As we shall see, other forms of attacks also arise that should be ruled out for a feasible model.
The ICE framework. Let us call an input \(({{hk}},x)\), consisting of the hash key \({{hk}} \) and a domain point x, to a hash function a full input. One way to view UCEs is that they adopt the indistinguishabilitybased approach (1) above, but restrict hash queries so that full inputs remain hidden from the attacker(s). It is clear that such hidden queries are not meaningful in the presence of a single adversary—any adversary knows its own queries—and hence UCEs come with two adversaries. Unpredictability together with denial of oracle access to \(D \) ensures that the x components of full inputs remain hidden from \(D \). On the other hand, the \({{hk}} \) components of full inputs remain hidden from \(S \) as the source is denied access to \({{hk}} \) (and no communication from \(D \) to \(S \) is allowed). As a result, full inputs \(({{hk}},x)\) remain hidden from both parties involved in a UCE attack.
This perspective allows us to build on UCEs and extend them as follows. In our new framework, which we call Interactive Computational Extractors (ICEs),^{3} a general mechanism for the joint generation of full inputs is enabled and adversarial restrictions that formalize what it means for full hash inputs to have high entropy are imposed.
We let two distinguishers \((D_1,D_2)\) to take part in an attack, and allow them to communicate via a bidirectional channel. Both distinguishers get access to a challenge hash oracle, which depending on a challenge bit implements either the real hash function or a (keyed) random oracle. To enable the two parties to make hidden queries, we introduce a shared writeonly tape that both \(D_1\) and \(D_2\) can write onto. When a distinguisher queries the hash oracle, the (real or ideal) hash of the full contents of the tape is returned. In contrast to UCEs, \(D_1\) or \(D_2\) can generate a hash key and perhaps modify it throughout the attack. This attack scenario is symmetric for \(D_1\) and \(D_2\) and, without loss of generality, the game terminates by \(D_2\) outputting with a bit. (Our formal definition, however, comes with a slightly more general return statement.) For a class \(\mathcal {C} \) of distinguishers, we define \(\mathrm {ICE} [\mathcal {C} ]\) security by demanding that the probability of guessing the challenge bit for any \(D=(D_1,D_2)\in \mathcal {C} \) is negligibly close to 1 / 2. See Fig. 1 (right) for a summary of this interaction.
Entropic queries. Similarly to \(\mathrm {UCE}\)s, the \(\mathrm {ICE}\) notion cannot be achieved without constraining the way the two distinguishers communicate. The main restriction that we introduce is analogous to statistical unpredictability for \(\mathrm {UCE}\)s: we demand the statistical unpredictability of full inputs to the hash function, including the hash key \({{hk}} \), from each distinguisher’s point of view. We choose a statistical, rather than a computational, notion so that our definitions do not become subject to the interactive versions of the attacks highlighted in [BFM14].^{4} More precisely, we require that when the hash oracle implements a keyed random oracle, no (possibly unbounded) predictor can guess a full input \(({{hk}},x)\) used to compute a hash value when it is provided with a distinguisher’s view consisting of its inputs, random coins, and all incoming messages and oracle responses.
Since our framework allows oracle access to both parties, unlike UCEs the two distinguishers can implicitly communicate via hash patterns as follows. Suppose \(D_2\) wants to leak a bit d to \(D_1\). Algorithm \(D_2\) starts by writing a random string onto the second half of the input and hands over the attack to \(D_1\). Algorithm \(D_1\) writes a random value to the first half of the input, calls \(\textsc {Hash} \) to receive a first hash value \(h_1\), and hands over the attack back to \(D_2\). Now algorithm \(D_2\), according to the value of d, either modifies the contents of the second half of the input tape or leaves them unchanged. \(D_1\) can recover d by obtaining a second hash value \(h_2\) and checking if \((h_1=h_2)\). The two distinguishers can also communicate via a bitfixing attack: \(D_2\) samples many (unpredictable) random values x conditioned on its hash value beginning with bit d, which \(D_1\) can then recover via a hash query.
In our unpredictability definition the predictor gets to see all hash responses, and hence if there are any repetitions they will be seen by the predictor. Unpredictability will therefore ensure that such repetition patterns will not leak any of the queries. Sometimes, however, we need to explicitly disallow any repeat queries to enable a security proof to go through. In such a scenario, we can ensure that there is no leakage via hash patterns either. Repeatfreeness appears in other related settings such as relatedkey attacks or correlatedinput hashing [BK03, GOR11].
1.4 Applications
BHK [BHK13a] use \(\mathrm {UCE}\)s to show that the encryption scheme of Black, Rogaway, and Shrimpton (BRS) [BRS03] is secure under relatedkey attacks (RKAs) and keydependentmessage (KDM) attacks as long as the related keys/keydependent messages are derived nonadaptively at the onset and without access to the hash key or previous ciphertexts.^{5}
As we shall see, ICE encompasses UCE as a special case, and the BRS scheme can also be instantiated under the above models using ICEs. We can however also obtain feasibility results that are outside the reach of \(\mathrm {UCE}\)s. A practically relevant and desirable level of RKA security is that corresponding to key offsets (the socalled xorRKA security [LRW02, BK09]). We show that \(\mathrm {ICE} \)s are sufficient to prove the full xorRKA security of the BRS scheme. Our formal result is more general and applies to the larger class of split functions that take the form \(\phi (K_1\Vert K_2) = \phi _1(K_1) \Vert \phi _2(K_2)\). (Such functions have been used to build RKAsecure PRFs [BC10], and also appear in other related contexts [CG14, LL12].) In addition to achieving stronger security guarantees, ICEs allow instantiating the BRS scheme using unkeyed hash functions, which is arguably closer to the original formulation of BRS.^{6}
We also strengthen the attainable KDM security guarantees for BRS by showing that adversaries can choose keydependent messages adaptively based on the hash key and also semiadaptively depending on previous ciphertexts. We prove that \(\mathrm {ICE} \)s are adaptively correlatedinput secure [GOR11] and that they relate well to other standard security properties of random oracles, such as pseudorandomness, randomness extraction, and oneway security (see full version). We leave it as open questions to see if full RKA beyond xor offsets or full KDM security can be established using extractorlike notaions.
1.5 Instantiations
BHK show that random oracles fulfill their strongest proposed \(\mathrm {UCE}\) notion, namely UCE security with respect to computationally unpredictable sources.^{7} We prove that random oracles are also \(\mathrm {ICE}\) secure. The significance of these results are twofold [BHK13a]: (1) there are no generic attacks on \(\mathrm {ICE}\)s and the model is structurally sound; and (2) a layered approach to security analysis can be enabled, whereby one first proves the security of a scheme under an \(\mathrm {ICE}\) assumption and then applies the RO model feasibility result. The latter is akin to security analyses carried out in the generic group model.
Practical hash functions, however, are not monolithic objects and often follow an iterative procedure to convert a fixedinputlength random oracle (FILRO) into a variableinputlength random oracle (VILRO). This, in turn, raises the question whether or not the above result can be brought closer to practice by demonstrating positive feasibility results for VILICEs in the FILRO model. A seemingly immediate way to establish this result would be to start with a hash function that is known to be indifferentiable from a VIL RO (e.g., the \(\mathsf {HMAC}\) or the \(\mathsf {NMAC}\) construction), and then apply the RO feasibility result above to conclude. This argument, however, fails as the \(\mathrm {ICE} \) game is multistaged and indifferentiability does not necessarily guarantee composition in such settings [RSS11].
Motivated by the above observations, we show both positive and negative feasibility results for ICEs. On the negative side, we show that the indifferentiable \(\mathsf {HMAC}\) and \(\mathsf {NMAC}\) constructions are provably \(\mathrm {ICE}\) insecure in the FILRO model. On the positive side, and building on Mittelbach’s techniques [Mit14], we prove that a keyed version of Liskov’s Zipper Hash [Lis07] is ICE secure (as a VIL hash function) under the assumption that the underlying compression function is a FILRO. Zipper Hash can be seen as a variant of the classical Merkle–Damgård [Dam90, Mer90] construction where the message blocks are processed twice in the forward and backward directions. Hence our results strengthen the VILRO feasibility result above, and also provide formal evidence for the (intuitive) added security guarantees that multipass hash functions seem to offer over their singlepass counterparts. For instance, combined with our RKA and KDM results, we may conclude that Zipper Hash can be safely used within the BRS scheme with no adverse affects on its security.
The above analysis can be further strengthened in at least two directions. First, one can weaken the underlying assumption and assume that the compression function underlying Zipper Hash is only a FILICE (rather than a FILRO). To this end, BHK [BHK14] give domain extenders for UCEs. Second, and motivated by the standardmodel realizations of ICEs and UCEs, we ask if these primitives can be based on plausible hardness assumptions. Brzuska and Mittelbach [BM14a, BM15] have recently shown positive results for UCEs with respect to restricted classes of sources.
2 Notation
We denote the security parameter by \(\lambda \in \mathbb {N}\), which is implicitly given to all algorithms (if not explicitly stated so) in the unary representation \({1^{\lambda }}\). By \({\{0,1\}}^\ell \) we denote the set of all bit strings of length \(\ell \) and \({\{0,1\}}^*\) is the set of all finitelength bit strings. For \(x,y\in {\{0,1\}}^*\) we denote their concatenation by \(x\Vert y\), the length of x by \(\leftx \right\), the ith bit of x by x[i], and the substring of x formed using bits i to j by x[i..j]. We denote the empty string by \(\varepsilon \). For X a finite set, \(\leftX \right\) denotes its cardinality, and Open image in new window denotes the action of sampling x uniformly at random from X. If \(\mathsf {Q}\) is a list and x a string then \(\mathsf {Q}: x\) denotes the list obtained by appending x to \(\mathsf {Q}\). Similarly, If \(\mathsf {Q}_1\) and \(\mathsf {Q}_2\) are lists, then \(\mathsf {Q}_1 : \mathsf {Q}_2\) denotes the concatenated list. Unless stated otherwise, algorithms are assumed to be randomized. We call an algorithm efficient or PPT if it runs in time polynomial in the security parameter. By \(y \leftarrow \mathcal {A}(x;r)\) we denote that y was output by algorithm \(\mathcal {A}\) on input x and randomness r. If \(\mathcal {A}\) is randomized and no randomness is specified, then we assume that \(\mathcal {A}\) is run with freshly sampled uniform random coins, and write Open image in new window . We use \(\mathsf {Coins}[A]\) to denote the polynomially long string of random coins r used by a PPT machine \(A\). We say a function \(\mathrm {negl}{\left( \lambda \right) }\) is negligible if \(\mathrm {negl}{\left( \lambda \right) } \in \lambda ^{\omega (1)}\).
Hash functions. In the line with [BHK13a], we consider the following (simplified) formalization of hash functions. A hash function consists of five PPT algorithms \(\mathsf {H}:=(\mathsf {H}.\mathsf {Kg} \), \(\mathsf {H}.\mathsf {Ev} \), \(\mathsf {H}.\mathsf {kl}\), \(\mathsf {H}.\mathsf {il}\), \(\mathsf {H}.\mathsf {ol})\) as follows. The keygeneration algorithm \(\mathsf {H}.\mathsf {Kg} \) gets the security parameter \({1^{\lambda }}\) as input and outputs a key \({{hk}} \in {\{0,1\}}^{\mathsf {H}.\mathsf {kl}(\lambda )}\), where \(\mathsf {H}.\mathsf {kl}(\lambda )\) is the keylength function. Algorithm \(\mathsf {H}.\mathsf {il}(\lambda )\) outputs the length of admissible inputs, which could take the special value \(*\) denoting the variablelength input space \({\{0,1\}}^*\). Algorithm \(\mathsf {H}.\mathsf {ol}(\lambda )\) outputs the length of admissible outputs, which we assumed to be a fixed polynomial function of the security parameter. The deterministic evaluation algorithm \(\mathsf {H}.\mathsf {Ev} \) takes as input the security parameter \({1^{\lambda }}\), a key \({{hk}} \), a point \(x\in {\{0,1\}}^{\mathsf {H}.\mathsf {il}(\lambda )}\), and generates a hash value \(\mathsf {H}.\mathsf {Ev} ({1^{\lambda }},{{hk}},x)\in {\{0,1\}}^{\mathsf {H}.\mathsf {ol}(\lambda )}\). To ease notation, we often suppress the security parameter and simply write \(\mathsf {H}.\mathsf {Ev} ({{hk}},x)\).
3 The \(\mathrm {ICE}\) Framework
We require \((D_1,D_2)\) not to leave any superfluous blank spaces on the joint tape. That is, a \(\textsc {Write} \) call must ensure that before the \(\textsc {Hash} \) oracle is called there do not exist indices \(i<j\) such that \(x[i]=\varepsilon \ne x[j]\) or \({{hk}} [i]=\varepsilon \ne {{hk}} [j]\). We also demand that the full inputs \(({{hk}},x)\) are valid in the sense that prior to a \(\textsc {Hash} \) call \({{hk}} \in {\{0,1\}}^{\mathsf {H}.\mathsf {kl}(\lambda )}\) and \(x \in {\{0,1\}}^{\mathsf {H}.\mathsf {il}(\lambda )}\). Although the distinguishers \(D_1\) and \(D_2\) are in general stateful algorithms, we omit the explicit handling of state values from the inputs and outputs of \(D _i\).
Class  Description 

\(\mathcal {C} _i^{w,q,r}\)  \(D_i\) places at most \(w(\lambda )\) queries to \(\textsc {Write} \), at most \(q(\lambda )\) queries to \(\textsc {Hash} \), and terminates after at most \(r(\lambda )\) invocations 
\(\mathcal {C} _i^\mathrm {poly}\)  \(D_i\) makes polynomially many oracle queries 
\(\mathcal {C} _i^\mathrm {ppt} \)  \(D_i\) runs in polynomial time on each invocation and terminates after a polynomial number of rounds 
\(\mathcal {C} _i^{0}\)  \(D_i\) sets \(b_i :=0\) in all invocations 
\(\mathcal {C} _i^\varepsilon \)  \(D_i\) sets \(L_{3i} :=\varepsilon \) in all invocations 
\(\mathcal {C} _i^{\mathrm {0}{\text {}}\mathrm {hk}} \)  \(D_i\) never writes onto the \({{hk}} \) part of the tape 
\(\mathcal {C} _i^{\mathrm {1}{\text {}}\mathrm {hk}} \)  On its first invocation, \(D_i\) writes a random \({{hk}} \) onto the \({{hk}} \)part of the tape. In subsequent invocations, \(D_i\) never writes onto the \({{hk}} \)part of the tape 
\(\mathcal {C} _i^\mathrm {dist}\)  \(D_i\) makes distinct queries to \(\textsc {Hash} \). That is, for lists \(\mathsf {Q}_1\) and \(\mathsf {Q}_2\) defined in Fig. 3, the combined list \(\mathsf {Q}_1 : \mathsf {Q}_2\) is repetitionfree. Note that \(\mathcal {C} _i^\mathrm {dist}= \mathcal {C} _{3i}^\mathrm {dist}= \mathcal {C} ^\mathrm {dist}\) 
\(\mathcal {C} _i^\mathrm {sup}\)  The probability that any (possibly unbounded) predictor \(P\) can guess a full query of \(D_i\) is negligible. We call this the class of statistically unpredictable \(D_i\). See Fig. 3 for the formal definition. Class \(\mathcal {C} _i^\mathrm {cup}\) is the computational analogue, where \(P\) is restricted to be ppt 
Note that the predicator only gets to see the hash responses for distinguisher \(D_i\)—these are within \(D_i\)’s view—and has to guess a query made by either distinguisher in the concatenated list \(\mathsf {Q}_1 : \mathsf {Q}_2\). It is easy to check that UCE security with respect to statistically unpredictable sources is equivalent to \(\mathrm {ICE} [\mathcal {C} ^{\mathrm {uce}} \cap \mathcal {C} ^\mathrm {sup}]\) security.
Remark. Since predictor \(P\) receives the full view of a distinguisher \(D_i\), it can perfectly simulate a run of \(D_i\) in the \(\mathrm {ICE} \) game with respect to a random implementation of the hash oracle, without any need to see the view of the partner distinguisher \(D_{3i}\). We will rely on this observation in our proofs.
4 Example Applications
Distinguisher classes used (above) and shown feasibility for (below). Here \(\mathcal {C} ^* := \mathcal {C} ^\mathrm {ppt} \cap \mathcal {C} ^\mathrm {dist}\cap \mathcal {C} _1^{\mathrm {1}{\text {}}\mathrm {hk}} \cap \mathcal {C} _2^{\mathrm {0}{\text {}}\mathrm {hk}} \cap \mathcal {C} _2^\varepsilon \).
Goal/Model  Class used/Achieved 

Split RKA  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _2^0\) 
Split KDM  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _1^0\) 
Split/clawfree CIH  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _2^0\) 
Extractor  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _1^0 \cap \mathcal {C} ^\varepsilon \cap \mathcal {C} ^{1,1,2}\) 
Weak PRF  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _1^0 \cap \mathcal {C} ^\varepsilon \cap \mathcal {C} ^{\mathrm {poly},\mathrm {poly},1}\) 
polyregular OWF  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} _1^0 \cap \mathcal {C} ^{1,1,1}\) 
VILROM  \(\mathcal {C} ^\mathrm {ppt} \cap \mathcal {C} ^\mathrm {cup}\) and \(\mathcal {C} ^\mathrm {poly}\cap \mathcal {C} ^\mathrm {sup}\); both contain \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\) 
FILROM  \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {cup}\), which contains \(\mathcal {C} ^* \cap \mathcal {C} ^\mathrm {sup}\) 
4.1 Split RKA Security
We show that the symmetric encryption scheme proposed by Black, Rogaway, and Shrimpton (BRS) [BRS03] is secure against relatedkey attacks (RKAs) when instantiated with an ICEsecure hash function. The encryption algorithm of the BRS scheme is implemented via \(\mathsf {Enc}^\mathsf {H}(K,M;R) := \left( R, M \oplus \mathsf {H}(K \Vert R) \right) \), for a hash function \(\mathsf {H}\), randomness R and key K. Recall that in an RKA, an adversary can obtain encryptions of messages of its choice under correlated keys (e.g., under K and \(K \oplus 1\)).
Split relatedkey derivation (RKD) functions \(\phi \) decompose into two subRKD functions \(\phi _1\) and \(\phi _2\) that are applied in parallel to two (fixed) substrings of the key: \(\phi (K_1\Vert K_2) = \phi _1(K_1) \Vert \phi _2(K_2)\).^{9} Split functions capture many RKA cases of interest including the case of xoring constants into keys. Without the minimal assumption that \(\phi \)’s have unpredictable outputs (i.e., the guessing probability of the outputs of \(\phi (K)\) over randomly chosen K is negligible) RKA security is not achievable [BK03]. In our proof, we will require a slightly stronger condition that the subRKD functions \(\phi _1(K_1)\) and \(\phi _2(K_2)\) are individually unpredictable. Note that offsetting keys via xor enjoys this property as xor induces a permutation over the two halves of the key.
The ICE adversary. Given an RKA adversary \(A\), we construct an \(\mathrm {ICE} \) adversary \((D_1,D_2)\), where \(D_1\) handles the left components of \(A\)’s RKA queries and \(D_2\) handles the right components as follows.

\(D_1(L_1)\): On initial invocation, generate a hash key \({{hk}} \), a random \(K_1\), and a random bit b. Store these values and write \({{hk}} \) onto the \({{hk}} \)part of the tape. Run \(A({{hk}})\) to get an RKA query \(((\phi _1,\phi _2),M_0,M_1)\). Output \((b_1,L_2) := (\bot ,\phi _2)\). Proceed as follows in subsequent invocations. Generate and store a random R and write \(\phi _1(K_1)\) onto the 1st segment (out of three segments) of the xpart of the tape and R onto its 3rd segment. Query \(\textsc {Hash} \) to get H. Recover R and resume \(A\) on \((R, H \oplus M_b)\) to get a new RKA query \(((\phi _1,\phi _2),M_0,M_1)\), or a bit \(b'\). If \(A\) outputs a bit \(b'\), return \((b_1,L_2) := (b=b',\varepsilon )\) and terminate. Else output \((b_1,L_2) := (\bot ,\phi _2)\).

\(D_2(L_2)\): When initially invoked, generate a random \(K_2\) and store it. In all invocations (including the first), recover \(\phi _2\) from \(L_2\). If \(\phi _2=\varepsilon \), return \((b_2,L_1) := (0,\varepsilon )\) and terminate. Else write \(\phi _2(K_2)\) onto the 2nd segment of the xpart of the tape. Output \((b_2,L_1) := (0,\varepsilon )\).
4.2 KDM Security
When the random oracle in the BRS scheme is instantiated with an \(\mathrm {ICE} \)secure hash function, we are able to show that the BRS scheme resists a partially adaptive form of KDM security for split keydependentmessage derivation (KDMD) functions \(\phi \). As for RKD functions, such KDMD functions consist of subKDMD functions \(\phi _1\) and \(\phi _2\) of the form \(\phi (K_1\Vert K_2) := \phi _1(K_1) \Vert \phi _2(K_2)\). The adaptivity level that we can tolerate is as follows. In an initial phase of the attack, the adversary can fully adaptively query split KDMD functions that do not depend on \(K_2\). That is, for these functions \(\phi _2(K_2)\) is constant and independent of \(K_2\) and its value can be predicted. In a second phase of the attack, the adversary can query split KDMD functions of its choice as long as either \(\phi _1(K_1)\) is constant or \(\phi _1(K_1)\) was used in the first phase. (We emphasize that these functions are not required to be unpredictable.) This model is strong enough to imply INDCPA security (without any restrictions), a case that could not be treated using UCEs.
The ICE adversary. Let \(A\) be a KDM adversary against the BRS scheme in the model above. Our \(\mathrm {ICE} [\mathcal {C} ^\mathrm {ppt} \cap \mathcal {C} ^\mathrm {sup}\cap \mathcal {C} ^\mathrm {dist}]\) adversary corresponding to \(A\) is as follows, where for simplicity we have assumed the lengths of keys, randomness and messages are all \(\ell \). (The ICE class can be further restricted as is shown in Table 1.) In this reduction, \(D_1\) faithfully runs the first stage of the attack, while \(D_2\) runs its second stage. To answer KDM queries, \(D_2\) relies on the “homomorphic” property that \(H \oplus (x_1\Vert x_2) = H \oplus (x_1\Vert 0^{x_2}) \oplus (0^{x_1}\Vert x_2)\).

\(D_1(L_1)\): When initially invoked, generate a random \({{hk}} \), \(K_1\) and b and store them. Write \({{hk}} \) to the \({{hk}} \)part and \(K_1\) to the 1st (out of three) segments of the xpart of the tape. (The segments are of lengths \(\ell /2\), \(\ell /2\) and \(\ell \) corresponding to \(K_1\), \(K_2\) and R respectively.) Output \((\bot ,\varepsilon )\). On the second invocation, run \(A({{hk}})\) and answer its KDM queries \(((\phi ^0_1,\phi ^0_2),(\phi ^1_1,\phi ^1_2))\) as follows. Write a fresh random value R onto the 3rd segment of the xpart of the tape. Call \(\textsc {Hash} \) to get H, and resume \(A\) on \((R, H \oplus (\phi _1(K_1)\Vert M_2^*))\), where \(M_2^*:=\phi _2(0^{\ell /2})\) is the right \(K_2\)independent part of the message. Continue this process until \(A\) decides to proceed to its second stage. Let \({st}_{A}\) denote \({A}\)’s state. Generate sufficiently many copies \((R_1,C'_1),\ldots ,(R_q,C'_q)\) of each of the KDM queries made in the first phase. Let \(\mathsf {List}_1\) denote the corresponding list of queried \(\phi ^b_1\). Return \((0,(b,{st}_A, (R_1,C'_1),\ldots ,(R_q,C'_q),\mathsf {List}_1))\) and terminate.

\(D_2(L_2)\): When initially invoked, generate a random \(K_2\), store it, and write it to the 2nd segment of the xpart of the tape. Hand the attack back to \(D_1\), by outputting \((\bot ,\varepsilon )\). On the second invocation, parse \(L_2\) appropriately as above. Resume \(A\) on \({st}_A\) and answer its KDM queries \(((\phi ^0_1,\phi ^0_2),(\phi ^1_1,\phi ^1_2))\) as follows. If \(\phi ^b_1 \in \mathsf {List}_1\) pick a fresh ciphertext \((R,C')\) corresponding to \(\phi ^b_1\) and complete the ciphertext preparation by setting \(C \leftarrow C \oplus (0^{\ell /2} \Vert \phi ^b_2(K_2))\). Otherwise generate a random R, write it onto the 3rd segment of the xpart of the tape, query \(\textsc {Hash} \) to get H, and set \(C \leftarrow H \oplus (\phi ^b_1(0^{\ell /2})\Vert \phi ^b_2(K_2))\). Resume \(A\) on \((R,C;{st}_A)\) and continue in this manner until \(A\) outputs a bit \(b'\). Return \((b=b',\varepsilon )\) and terminate.
5 Feasibility
In this section we start by showing that random oracles are \(\mathrm {ICE}\) secure with respect to interesting distinguisher classes (in particular, with respect to the restrictions needed for the presented applications). We then consider the ICE security of practical hash constructions built from fixinputlength (FIL) ROs. In particular, we look at a keyed variant of Liskov’s Zipper Hash [Lis07] and show that it achieves \(\mathrm {ICE} \) security in the FILRO model. Interestingly, we show that both \(\mathsf {HMAC}\) and \(\mathsf {NMAC}\) constructions [BCK96], which were recently shown to be \(\mathrm {UCE}\) secure in FILROM [Mit14], fail to be \(\mathrm {ICE}\) secure. This result yields a natural counterexample to the composability of \(\mathsf {HMAC}\) in multistage settings, similarly to that given by Ristenpart, Shacham, and Shrimpton in [RSS11]. Furthermore, it provides a separation between ICE and UCE. Our results also demonstrate that Zipper Hash can provide a higher level of security compared to \(\mathsf {HMAC}\) when used in multistage settings.
5.1 ICEs from Random Oracles
BHK [BHK13b] show that UCEsecure hash functions can be provably constructed in the RO model. The philosophical justifications of this result are that there are no structural weaknesses in the definitional framework, and more importantly, a layered approach to protocol design in the RO model can be enabled [BHK13b]. We show that \(\mathrm {ICE} \)s also enjoy RO feasibility.
Let \(\mathsf {H}.\mathsf {kl}(\cdot )\) and \(\mathsf {H}.\mathsf {ol}(\cdot )\) be two arbitrary functions as in the syntax of a hash function. Let \(\mathcal {R} \) be a family of variableinputlength (VIL) ROs (i.e., with domain \({\{0,1\}}^*\)) and range \({\{0,1\}}^{\mathsf {H}.\mathsf {ol}(\lambda )}\). We construct the required hash function \(\mathsf {H} ^\mathcal {R} \) by defining the keygeneration algorithm \(\mathsf {H}.\mathsf {Kg} ({1^{\lambda }})\) to return a random Open image in new window and the evaluation algorithm \(\mathsf {H}.\mathsf {Ev} ^\mathcal {R} ({{hk}},x)\) to return \(\mathcal {R} ({{hk}} \Vert x)\). Our first feasibility result is as follows.
Theorem 1
5.2 VILICEs from Ideal Compression
Practical variableinputlength (VIL) hash functions are not monolithic objects. They often follow iterative modes of chaining that convert a fixinputlength (FIL) compression function to one that accepts variablelength inputs. This design principle has been successfully validated via the indifferentiability framework of Maurer et al. [MRH04, CDMP05], whereby an indifferentiable hashfunction construction is shown to securely compose when used in place of a random oracle. As pointed out in [RSS11], the indifferentiability framework only guarantees composition in singlestage environments. The ICE and UCE games, however, are inherently multistaged and lie outside the reach of (plain) indifferentiability. Mittelbach [Mit14] develops new techniques to extend the reach of (plain) indifferentiability to certain classes of multistage games. In particular, he shows that the \(\mathsf {HMAC}\) and \(\mathsf {NMAC}\) constructions are UCE secure. Interestingly, we show that these results do not carry over to the \(\mathrm {ICE} \) model: \(\mathsf {HMAC}\) and \(\mathsf {NMAC}\) provably fail to be ICE secure. On the other hand, we build on Mittelbach’s techniques to prove that a variant of Zipper Hash [Lis07] is provably ICE secure.
Zipper Hash. The above attack raises the question if any iterative hash function can be \(\mathrm {ICE} [\mathcal {C} ]\) secure for a meaningful class of distinguishers \(\mathcal {C} \). We show that a hybrid construction of a keyed version of Liskov’s Zipper Hash construction [Lis07] and chopped Merkle–Damgård (chopMD) of Coron et al. [CDMP05] is ICE secure. Zipper Hash can be regarded as a basic Merkle–Damgård scheme where the message is processed twice, the second time in reversed block order. chopMD refers to the construction where a hash value consists only of the first half of the output bits of the final compression function. Our hybrid construction results from adding the chop step to Zipper Hash. Furthermore, we consider a keyed variant of Zipper Hash by prepending the hash key to the message. We assume that key length matches block length, which means that the first and last evaluations of the compression function operate on the hash key. We denote this keyed variant of Zipper Hash by \(\mathsf {chop{{\text {}}}KZIP}\). Figure 5 shows a schematic diagram of the construction.
Theorem 2
Note that class \(\mathcal {C} \) above contains that class used to attack \(\mathsf {HMAC}\) and hence \(\mathsf {chop{{\text {}}}KZIP}^\mathsf {h}\) provably achieves a higher level of security in multistage games. We note that the reach of the above feasibility result includes all applications scenarios listed in Table 1. In particular, \(\mathsf {chop{{\text {}}}KZIP}^\mathsf {h}\) can security replace the random oracle in these applications. For this also note that we can easily drop \(\mathcal {C} _1^0\) by requesting that in the last round \(D _1\) outputs a guess for b which \(D _2\) echoes. With the other restrictions present this change is without loss of generality.
This result cannot be strengthened for the (large) adversarial classes that were used in Theorem 1. To see this, consider two distinguishers that engage in a distributed computation of \(\mathsf {chop{{\text {}}}KZIP}^\mathsf {h}\) hash values as follows. Distinguisher \(D_1\) knows \({{hk}} \) and \(m_1\) and \(D_2\) knows \(m_2\), where message \(m := m_1\Vert m_2\) is being hashed. Distinguisher \(D_1\) computes an intermediate hash digest using \(({{hk}},m_1)\) and forwards it to \(D_2\). Distinguisher \(D_2\) now computes another iteration of the hash using \(m_2\) and forwards the result to \(D_1\). Distinguisher \(D_1\) can now complete the hash computation using its knowledge of \(({{hk}},m_1)\) and the intermediate hash digest that it receives.
A straightforward generalization of this attack also rules out multipass variants of \(\mathsf {chop{{\text {}}}KZIP}^\mathsf {h}\) (where messages are processed multiple times in the forward and backward directions), including those whose number of passes is not fixed a priori and can depend on the number of message blocks. This is due to the fact that the number of rounds in an \(\mathrm {ICE} \) attack is not fixed. This, in turn, raises the question if \(\mathrm {ICE} [\mathcal {C} ^\mathrm {poly}\cap \mathcal {C} ^\mathrm {sup}]\) is feasible in the FILRO model. We conclude the paper with a candidate construction that we conjecture to reach this level of security.
Footnotes
 1.
Such computational \(\mathrm {UCE}\) notions are intrinsically needed for applications such as simultaneous hadcore bits and deterministic PKEs.
 2.
This can be viewed as an interactive analogue of BFM’s attack [BFM14].
 3.
In UCEs, “universal” refers to the fact that extraction should work with respect to universal (i.e., all admissible) sources. Analogously, “interactive” in ICEs refers to the fact that extraction should work for sources that can interact.
 4.
This is also motivated by impossibility results for statistically secure twoparty protocols.
 5.
Recall that in RKA security the adversary can see encryptions of messages under keys \(\phi (K)\) for a random K and functions \(\phi \) of its choice. In KDM security the adversary can see encryptions of \(\phi (K)\), under a random key K, for \(\phi \)’s of its choice.
 6.
BRS [BRS03] analyze their scheme in the unkeyed RO model, which translates to unkeyed instantiations in practice.
 7.
Note that this does not contradict the BFM attack as ROs do not have succinct descriptions.
 8.
We emphasize that computational notions are still valuable as combined with our feasibility results, they would enable easier and more modular security proofs in the RO model.
 9.
For simplicity we assume that these are just the left and right halves of the key. Our proof will however also apply to any two substrings of superlogarithmic lengths.
 10.
Indeed, \(\mathsf {MixHash}\) is a (highly) offline function: for \(\alpha \in [0,1]\), it requires space roughly \(n\ell \sqrt{1\alpha }\) bits after a fraction \(\alpha \) of the \(n\ell (\ell +1)/2\) bits are processed.
Notes
Acknowledgments
The authors would like to thank Christina Brzuska for taking part in the early stages of this work. Pooya Farshim was supported in part by grant ANR14CE280003 (Project EnBid).
References
 [BC10]Bellare, M., Cash, D.: Pseudorandom functions and permutations provably secure against relatedkey attacks. In: Rabin, T. (ed.) CRYPTO 2010. LNCS, vol. 6223, pp. 666–684. Springer, Heidelberg (2010)CrossRefGoogle Scholar
 [BCFW09]Boldyreva, A., Cash, D., Fischlin, M., Warinschi, B.: Foundations of nonmalleable hash and oneway functions. In: Matsui, M. (ed.) ASIACRYPT 2009. LNCS, vol. 5912, pp. 524–541. Springer, Heidelberg (2009)CrossRefGoogle Scholar
 [BCK96]Bellare, M., Canetti, R., Krawczyk, H.: Keying hash functions for message authentication. In: Koblitz, N. (ed.) CRYPTO 1996. LNCS, vol. 1109, pp. 1–15. Springer, Heidelberg (1996)Google Scholar
 [BFM14]Brzuska, C., Farshim, P., Mittelbach, A.: Indistinguishability obfuscation and UCEs: the case of computationally unpredictable sources. In: Garay, J.A., Gennaro, R. (eds.) CRYPTO 2014, Part I. LNCS, vol. 8616, pp. 188–205. Springer, Heidelberg (2014)CrossRefGoogle Scholar
 [BGI+01]Barak, B., Goldreich, O., Impagliazzo, R., Rudich, S., Sahai, A., Vadhan, S.P., Yang, K.: On the (Im)possibility of Obfuscating Programs. In: Kilian, J. (ed.) CRYPTO 2001. LNCS, vol. 2139, pp. 1–18. Springer, Heidelberg (2001)CrossRefGoogle Scholar
 [BH15]Bellare, M., Hoang, V.T.: Resisting randomness subversion: fast deterministic and hedged publickey encryption in the standard model. In: Oswald, E., Fischlin, M. (eds.) EUROCRYPT 2015. LNCS, vol. 9057, pp. 627–656. Springer, Heidelberg (2015)Google Scholar
 [BHK13a]Bellare, M., Hoang, V.T., Keelveedhi, S.: Instantiating random oracles via UCEs. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part II. LNCS, vol. 8043, pp. 398–415. Springer, Heidelberg (2013)CrossRefGoogle Scholar
 [BHK13b]Bellare, M., Hoang, V.T., Keelveedhi, S.: Instantiating random oracles via UCEs. Cryptology ePrint Archive, Report 2013/424 (2013). http://eprint.iacr.org/2013/424
 [BHK14]Bellare, M., Hoang, V.T., Keelveedhi, S.: Cryptography from compression functions: the UCE bridge to the ROM. In: Garay, J.A., Gennaro, R. (eds.) CRYPTO 2014, Part I. LNCS, vol. 8616, pp. 169–187. Springer, Heidelberg (2014)CrossRefGoogle Scholar
 [BK03]Bellare, M., Kohno, T.: A theoretical treatment of relatedkey attacks: RKAPRPs, RKAPRFs, and applications. In: Biham, E. (ed.) EUROCRYPT 2003. LNCS, vol. 2656, pp. 491–506. Springer, Heidelberg (2003)CrossRefGoogle Scholar
 [BK09]Biryukov, A., Khovratovich, D.: Relatedkey cryptanalysis of the full AES192 and AES256. In: Matsui, M. (ed.) ASIACRYPT 2009. LNCS, vol. 5912, pp. 1–18. Springer, Heidelberg (2009)CrossRefGoogle Scholar
 [BK15]Bellare, M., Keelveedhi, S.: Interactive messagelocked encryption and secure deduplication. In: Katz, J. (ed.) PKC 2015. LNCS, vol. 9020, pp. 516–538. Springer, Heidelberg (2015)Google Scholar
 [BM14a]Brzuska, C., Mittelbach, A.: Indistinguishability obfuscation versus multibit point obfuscation with auxiliary input. In: Sarkar, P., Iwata, T. (eds.) ASIACRYPT 2014, Part II. LNCS, vol. 8874, pp. 142–161. Springer, Heidelberg (2014)Google Scholar
 [BM14b]Brzuska, C., Mittelbach, A.: Using indistinguishability obfuscation via UCEs. In: Sarkar, P., Iwata, T. (eds.) ASIACRYPT 2014, Part II. LNCS, vol. 8874, pp. 122–141. Springer, Heidelberg (2014)Google Scholar
 [BM15]Brzuska, C., Mittelbach, A.: Universal computational extractors and the superfluous padding assumption for indistinguishability obfuscation. Cryptology ePrint Archive, Report 2015/581 (2015). http://eprint.iacr.org/
 [BR93]Bellare, M., Rogaway, P.: Random oracles are practical: a paradigm for designing efficient protocols. In: Ashby, V. (ed.) ACM CCS 1993, pp. 62–73. ACM Press, November 1993Google Scholar
 [BRS03]Black, J., Rogaway, P., Shrimpton, T.: Encryptionscheme security in the presence of keydependent messages. In: Nyberg, K., Heys, H.M. (eds.) SAC 2002. LNCS, vol. 2595, pp. 62–75. Springer, Heidelberg (2003)CrossRefGoogle Scholar
 [BST15]Bellare, M., Stepanovs, I., Tessaro, S.: Contention in cryptoland: obfuscation, leakage and UCE. Cryptology ePrint Archive, Report 2015/487 (2015). http://eprint.iacr.org/2015/487 Google Scholar
 [Can97]Canetti, R.: Towards realizing random oracles: hash functions that hide all partial information. In: Kaliski Jr., B.S. (ed.) CRYPTO 1997. LNCS, vol. 1294, pp. 455–469. Springer, Heidelberg (1997)CrossRefGoogle Scholar
 [CD08]Canetti, R., Dakdouk, R.R.: Extractable perfectly oneway functions. In: Aceto, L., Damgård, I., Goldberg, L.A., Halldórsson, M.M., Ingólfsdóttir, A., Walukiewicz, I. (eds.) ICALP 2008, Part II. LNCS, vol. 5126, pp. 449–460. Springer, Heidelberg (2008)CrossRefGoogle Scholar
 [CD09]Canetti, R., Dakdouk, R.R.: Towards a theory of extractable functions. In: Reingold, O. (ed.) TCC 2009. LNCS, vol. 5444, pp. 595–613. Springer, Heidelberg (2009)CrossRefGoogle Scholar
 [CDMP05]Coron, J.S., Dodis, Y., Malinaud, C., Puniya, P.: MerkleDamgård revisited: how to construct a hash function. In: Shoup, V. (ed.) CRYPTO 2005. LNCS, vol. 3621, pp. 430–448. Springer, Heidelberg (2005)CrossRefGoogle Scholar
 [CG14]Cheraghchi, M., Guruswami, V.: Nonmalleable coding against bitwise and splitstate tampering. In: Lindell, Y. (ed.) TCC 2014. LNCS, vol. 8349, pp. 440–464. Springer, Heidelberg (2014)CrossRefGoogle Scholar
 [CGH98]Canetti, R., Goldreich, O., Halevi, S.: The random oracle methodology, revisited (preliminary version). In: 30th ACM STOC, pp. 209–218. ACM Press, May 1998Google Scholar
 [Dam90]Damgård, I.B.: A design principle for hash functions. In: Brassard, G. (ed.) CRYPTO 1989. LNCS, vol. 435, pp. 416–427. Springer, Heidelberg (1990)Google Scholar
 [DGG+15]Dodis, Y., Ganesh, C., Golovnev, A., Juels, A., Ristenpart, T.: A formal treatment of backdoored pseudorandom generators. In: Oswald, E., Fischlin, M. (eds.) EUROCRYPT 2015. LNCS, vol. 9056, pp. 101–126. Springer, Heidelberg (2015)Google Scholar
 [GGH+13]Garg, S., Gentry, C., Halevi, S., Raykova, M., Sahai, A., Waters, B.: Candidate indistinguishability obfuscation and functional encryption for all circuits. In: 54th FOCS, pp. 40–49. IEEE Computer Society Press, October 2013Google Scholar
 [GOR11]Goyal, V., O’Neill, A., Rao, V.: Correlatedinput secure hash functions. In: Ishai, Y. (ed.) TCC 2011. LNCS, vol. 6597, pp. 182–200. Springer, Heidelberg (2011)CrossRefGoogle Scholar
 [Lis07]Liskov, M.: Constructing an ideal hash function from weak ideal compression functions. In: Biham, E., Youssef, A.M. (eds.) SAC 2006. LNCS, vol. 4356, pp. 358–375. Springer, Heidelberg (2007)CrossRefGoogle Scholar
 [LL12]Liu, F.H., Lysyanskaya, A.: Tamper and leakage resilience in the splitstate model. In: SafaviNaini, R., Canetti, R. (eds.) CRYPTO 2012. LNCS, vol. 7417, pp. 517–532. Springer, Heidelberg (2012)CrossRefGoogle Scholar
 [LRW02]Liskov, M., Rivest, R.L., Wagner, D.: Tweakable block ciphers. In: Yung, M. (ed.) CRYPTO 2002. LNCS, vol. 2442, pp. 31–46. Springer, Heidelberg (2002)CrossRefGoogle Scholar
 [Mer90]Merkle, R.C.: One way hash functions and DES. In: Brassard, G. (ed.) CRYPTO 1989. LNCS, vol. 435, pp. 428–446. Springer, Heidelberg (1990)Google Scholar
 [MH14]Matsuda, T., Hanaoka, G.: Chosen ciphertext security via UCE. In: Krawczyk, H. (ed.) PKC 2014. LNCS, vol. 8383, pp. 56–76. Springer, Heidelberg (2014)CrossRefGoogle Scholar
 [Mit14]Mittelbach, A.: Salvaging indifferentiability in a multistage setting. In: Nguyen, P.Q., Oswald, E. (eds.) EUROCRYPT 2014. LNCS, vol. 8441, pp. 603–621. Springer, Heidelberg (2014)CrossRefGoogle Scholar
 [MRH04]Maurer, U.M., Renner, R.S., Holenstein, C.: Indifferentiability, impossibility results on reductions, and applications to the random oracle methodology. In: Naor, M. (ed.) TCC 2004. LNCS, vol. 2951, pp. 21–39. Springer, Heidelberg (2004)CrossRefGoogle Scholar
 [Nie02]Nielsen, J.B.: Separating random oracle proofs from complexity theoretic proofs: the noncommitting encryption case. In: Yung, M. (ed.) CRYPTO 2002. LNCS, vol. 2442, pp. 111–126. Springer, Heidelberg (2002)CrossRefGoogle Scholar
 [RSS11]Ristenpart, T., Shacham, H., Shrimpton, T.: Careful with composition: limitations of the indifferentiability framework. In: Paterson, K.G. (ed.) EUROCRYPT 2011. LNCS, vol. 6632, pp. 487–506. Springer, Heidelberg (2011)CrossRefGoogle Scholar