Total of 170 points

  1. (50 points) Here is one possible security definition for a witness encryption scheme: it is composed of two efficient algorithms \((E,D)\) with the following property. \(E\) is a probabilistic algorithm that takes as input a circuit \(C:{\{0,1\}}^n\rightarrow{\{0,1\}}\) and a message \(b\in{\{0,1\}}\) and outputs \(c=E_C(b)\). \(D\) takes as input a string \(w\) and a ciphertext \(c\), and the condition we require is that if \(C(w)=1\) then \(D_w(E_C(b))=b\). The notion of security is that if there exists no \(w\) such that \(C(w)=1\) then the distributions \(E_C(0)\) and \(E_C(1)\) are computationally indistinguishable (the distributions are over the coins of the encryption algorithm).
  1. (25 points) Prove that under the PRG assumption, witness encryption implies a public key encryption scheme. See footnote for hint1
  1. (25 points) Give a construction of a witness encryption scheme using an indistinguishability obfuscator \({\mathcal{O}}\). See footnote for hint2
  1. (60 points) A puncturable PRF is a pseudorandom function collection \(\{ f_s \}\) such that for every input \(x^*\), there is a way to map an index \(s\) into an index \(s^*=PUNCTURE(s,x^*)\) that allows to compute the function \(f_s\) on every input except \(x^*\). That is, there is some efficient algorithm \(EVAL\) such that \(EVAL(s^*,x)=f_s(x)\) for every \(x \neq x^*\) but such that even given \(s^*\), the value \(f_s(x^*)\) is comptuationally indstinguishable from a uniform value in \({\{0,1\}}^n\).
  1. (30 points) Show that under the PRG assumption, there exists a puncturable PRF. See footnote for hint3
  1. (30 points) Suppose that \({\mathcal{O}}\) is an IO obfuscator, \(G:{\{0,1\}}^n\rightarrow{\{0,1\}}^{3n}\) is a PRG and that \(\{ f_s \}\) (where \(f_s:{\{0,1\}}^{|s|}\rightarrow{\{0,1\}}^{|s|}\) is a puncturable PRF. Prove that the following is a selectively secure digital signature scheme, where by this we mean a scheme that satisfies the relaxed definition where the attacker must declare the message \(m^*\) on which she will forge a signature at the beginning of the chosen-message-attack game, before seeing the public key.

As a first step, worth 15 points, for every \(m^*\), consider the following circuit \(V^*_{m^*,s^*,z}\): for \(m\neq m^*\) \(V^*_{m^*,s^*,z}(m,\sigma)\) outputs \(1\) iff \(G(EVAL(s^*,m))=G(\sigma)\) and for \(m=m^*\), \(V^*_{m^*,s^*,z}(m,\sigma)\) outputs \(1\) iff \(G(\sigma)=z\). Prove that if \(s^* = PUNCTURE(m^*)\) and \(z=G(f_s(m^*))\) then \(V^*_{m^*,s^*,z}\) computes the same function as \(V_s\). By padding you can assume they have the same size as well.

See footnote for a hint how to complete the proof4

  1. (60 points) Suppose that Bob wants Alice to compute for him a function \(f(x)\) that is polynomial time computable but still takes too much time for him to compute online (though he can invest this time in a preprocessing step, before he learns the input \(x\) he needs to compute it for). Consider the following protocols for doing so using an FHE \((G,E,D,EVAL)\). We will also assume \(EVAL\) is a deterministic function.

Protocol 1:

  1. (20 points) Prove that the protocol satisfies the following notion of security: for every efficient strategy \(A\) for Alice, either Bob rejects with probability at least \(1/3\) or Bob outputs the correct output with probability at least \(1/3\).

  2. (20 points) Suppose that we run Protocol 1 twice for two inputs \(x_1,x_2\) with the same preprocessing step. The notion of security is now that for every efficient strategy \(A\) for Alice, either Bob rejects with probability at least \(1/3\) or Bob outputs the correct outputs for both \(x_1\) and \(x_2\) (i.e., \(f(x_1)\) and \(f(x_2)\)) with probability at least \(1/3\). Prove that this protocol satisfies this notion of security or give a counterexample (a strategy for Alice that would violate this property).

  3. (20 points) Consider the following protocol:

Protocol 2:

Prove that for every polynomial \(k\) and \(x_1,\ldots,x_k\), Protocol 2 satisfies the property that if we run the processing step once and then run the protocol \(k\) times with inputs \(x_1,\ldots,x_k\) then for every efficient strategy of Alice, either Bob rejects with probability at least \(1/3\), or he outputs all the correct \(k\) outputs with probability at least \(1/3\).


  1. The public key can be a string \(y=G(w)\) where \(G:{\{0,1\}}^n\rightarrow{\{0,1\}}^{2n}\) is a PRG, and the private key can be \(w\).

  2. One can phrase the goal of the encryption algorithm in a witness encryption scheme as transforming the circuit \(C\) and message \(b\) to some \(C'\) that maps \(w\) to \(b\) if \(C(w)=1\) and maps \(w\) to error (that can be encoded in some for, e.g., as \(0\)) if \(C(w)=0\). Of course one needs to ensure that it won’t be possible to extract \(b\) from \(C'\) if there is no \(w\) satisfying \(C(w)=1\).

  3. hint3

  4. Think of the following series of hybrids. First we can modify the key from the obfuscation of \(V_s\) to the obfuscation of \(V_{m^*,s^*,G(f_s(m^*))}\) and claim that the attackers success probability will stay the same due to the security of the IO scheme. Then we can transform the last output to \(G(U_n)\) and claim that there the success would still be the same due to the punctured PRF security. Finally we can modify the value \(G(U_n)\) to \(U_{3n}\) and claim that the suceess should still be the same due to the security of the PRG. But at this point, eith very high probability the verification algorithm \(V_{m^*,s^*,z}\) outputs \(0\) on every input of the form \((m^*,\sigma)\).