def example_bfv_basics(): print("Example: BFV Basics") #In this example, we demonstrate performing simple computations (a polynomial #evaluation) on encrypted integers using the BFV encryption scheme. # #The first task is to set up an instance of the EncryptionParameters class. #It is critical to understand how the different parameters behave, how they #affect the encryption scheme, performance, and the security level. There are #three encryption parameters that are necessary to set: # # - poly_modulus_degree (degree of polynomial modulus); # - coeff_modulus ([ciphertext] coefficient modulus); # - plain_modulus (plaintext modulus; only for the BFV scheme). # #The BFV scheme cannot perform arbitrary computations on encrypted data. #Instead, each ciphertext has a specific quantity called the `invariant noise #budget' -- or `noise budget' for short -- measured in bits. The noise budget #in a freshly encrypted ciphertext (initial noise budget) is determined by #the encryption parameters. Homomorphic operations consume the noise budget #at a rate also determined by the encryption parameters. In BFV the two basic #operations allowed on encrypted data are additions and multiplications, of #which additions can generally be thought of as being nearly free in terms of #noise budget consumption compared to multiplications. Since noise budget #consumption compounds in sequential multiplications, the most significant #factor in choosing appropriate encryption parameters is the multiplicative #depth of the arithmetic circuit that the user wants to evaluate on encrypted #data. Once the noise budget of a ciphertext reaches zero it becomes too #corrupted to be decrypted. Thus, it is essential to choose the parameters to #be large enough to support the desired computation; otherwise the result is #impossible to make sense of even with the secret key. parms = EncryptionParameters(scheme_type.BFV) #The first parameter we set is the degree of the `polynomial modulus'. This #must be a positive power of 2, representing the degree of a power-of-two #cyclotomic polynomial; it is not necessary to understand what this means. # #Larger poly_modulus_degree makes ciphertext sizes larger and all operations #slower, but enables more complicated encrypted computations. Recommended #values are 1024, 2048, 4096, 8192, 16384, 32768, but it is also possible #to go beyond this range. # #In this example we use a relatively small polynomial modulus. Anything #smaller than this will enable only very restricted encrypted computations. poly_modulus_degree = 4096 parms.set_poly_modulus_degree(poly_modulus_degree) #Next we set the [ciphertext] `coefficient modulus' (coeff_modulus). This #parameter is a large integer, which is a product of distinct prime numbers, #each up to 60 bits in size. It is represented as a vector of these prime #numbers, each represented by an instance of the SmallModulus class. The #bit-length of coeff_modulus means the sum of the bit-lengths of its prime #factors. # #A larger coeff_modulus implies a larger noise budget, hence more encrypted #computation capabilities. However, an upper bound for the total bit-length #of the coeff_modulus is determined by the poly_modulus_degree, as follows: # # +----------------------------------------------------+ # | poly_modulus_degree | max coeff_modulus bit-length | # +---------------------+------------------------------+ # | 1024 | 27 | # | 2048 | 54 | # | 4096 | 109 | # | 8192 | 218 | # | 16384 | 438 | # | 32768 | 881 | # +---------------------+------------------------------+ # #These numbers can also be found in native/src/seal/util/hestdparms.h encoded #in the function SEAL_HE_STD_PARMS_128_TC, and can also be obtained from the #function # # CoeffModulus::MaxBitCount(poly_modulus_degree). # #For example, if poly_modulus_degree is 4096, the coeff_modulus could consist #of three 36-bit primes (108 bits). # #Microsoft SEAL comes with helper functions for selecting the coeff_modulus. #For new users the easiest way is to simply use # # CoeffModulus::BFVDefault(poly_modulus_degree), # #which returns std::vector<SmallModulus> consisting of a generally good choice #for the given poly_modulus_degree. parms.set_coeff_modulus(CoeffModulus.BFVDefault(poly_modulus_degree)) #The plaintext modulus can be any positive integer, even though here we take #it to be a power of two. In fact, in many cases one might instead want it #to be a prime number; we will see this in later examples. The plaintext #modulus determines the size of the plaintext data type and the consumption #of noise budget in multiplications. Thus, it is essential to try to keep the #plaintext data type as small as possible for best performance. The noise #budget in a freshly encrypted ciphertext is # # ~ log2(coeff_modulus/plain_modulus) (bits) # #and the noise budget consumption in a homomorphic multiplication is of the #form log2(plain_modulus) + (other terms). # #The plaintext modulus is specific to the BFV scheme, and cannot be set when #using the CKKS scheme. parms.set_plain_modulus(1024) #Now that all parameters are set, we are ready to construct a SEALContext #object. This is a heavy class that checks the validity and properties of the #parameters we just set. context = SEALContext.Create(parms) #Print the parameters that we have chosen. print("Set encryption parameters and print") print_parameters(context) print("~~~~~~ A naive way to calculate 4(x^2+1)(x+1)^2. ~~~~~~") #The encryption schemes in Microsoft SEAL are public key encryption schemes. #For users unfamiliar with this terminology, a public key encryption scheme #has a separate public key for encrypting data, and a separate secret key for #decrypting data. This way multiple parties can encrypt data using the same #shared public key, but only the proper recipient of the data can decrypt it #with the secret key. # #We are now ready to generate the secret and public keys. For this purpose #we need an instance of the KeyGenerator class. Constructing a KeyGenerator #automatically generates the public and secret key, which can immediately be #read to local variables. keygen = KeyGenerator(context) public_key = keygen.public_key() secret_key = keygen.secret_key() #To be able to encrypt we need to construct an instance of Encryptor. Note #that the Encryptor only requires the public key, as expected. encryptor = Encryptor(context, public_key) #Computations on the ciphertexts are performed with the Evaluator class. In #a real use-case the Evaluator would not be constructed by the same party #that holds the secret key. evaluator = Evaluator(context) #We will of course want to decrypt our results to verify that everything worked, #so we need to also construct an instance of Decryptor. Note that the Decryptor #requires the secret key. decryptor = Decryptor(context, secret_key) #As an example, we evaluate the degree 4 polynomial # # 4x^4 + 8x^3 + 8x^2 + 8x + 4 # #over an encrypted x = 6. The coefficients of the polynomial can be considered #as plaintext inputs, as we will see below. The computation is done modulo the #plain_modulus 1024. # #While this examples is simple and easy to understand, it does not have much #practical value. In later examples we will demonstrate how to compute more #efficiently on encrypted integers and real or complex numbers. # #Plaintexts in the BFV scheme are polynomials of degree less than the degree #of the polynomial modulus, and coefficients integers modulo the plaintext #modulus. For readers with background in ring theory, the plaintext space is #the polynomial quotient ring Z_T[X]/(X^N + 1), where N is poly_modulus_degree #and T is plain_modulus. # #To get started, we create a plaintext containing the constant 6. For the #plaintext element we use a constructor that takes the desired polynomial as #a string with coefficients represented as hexadecimal numbers. x = 6 x_plain = Plaintext(str(x)) print("Express x = {} as a plaintext polynomial 0x{}.".format( x, x_plain.to_string())) #We then encrypt the plaintext, producing a ciphertext. x_encrypted = Ciphertext() print("Encrypt x_plain to x_encrypted.") encryptor.encrypt(x_plain, x_encrypted) #In Microsoft SEAL, a valid ciphertext consists of two or more polynomials #whose coefficients are integers modulo the product of the primes in the #coeff_modulus. The number of polynomials in a ciphertext is called its `size' #and is given by Ciphertext::size(). A freshly encrypted ciphertext always #has size 2. print(" + size of freshly encrypted x: {}".format(x_encrypted.size())) #There is plenty of noise budget left in this freshly encrypted ciphertext. print(" + noise budget in freshly encrypted x: {} bits".format( decryptor.invariant_noise_budget(x_encrypted))) #We decrypt the ciphertext and print the resulting plaintext in order to #demonstrate correctness of the encryption. x_decrypted = Plaintext() decryptor.decrypt(x_encrypted, x_decrypted) print(" + decryption of x_encrypted: 0x{} ...... Correct.".format( x_decrypted.to_string())) #When using Microsoft SEAL, it is typically advantageous to compute in a way #that minimizes the longest chain of sequential multiplications. In other #words, encrypted computations are best evaluated in a way that minimizes #the multiplicative depth of the computation, because the total noise budget #consumption is proportional to the multiplicative depth. For example, for #our example computation it is advantageous to factorize the polynomial as # # 4x^4 + 8x^3 + 8x^2 + 8x + 4 = 4(x + 1)^2 * (x^2 + 1) # #to obtain a simple depth 2 representation. Thus, we compute (x + 1)^2 and #(x^2 + 1) separately, before multiplying them, and multiplying by 4. # #First, we compute x^2 and add a plaintext "1". We can clearly see from the #print-out that multiplication has consumed a lot of noise budget. The user #can vary the plain_modulus parameter to see its effect on the rate of noise #budget consumption. print("Compute x_sq_plus_one (x^2+1).") x_sq_plus_one = Ciphertext() evaluator.square(x_encrypted, x_sq_plus_one) plain_one = Plaintext("1") evaluator.add_plain_inplace(x_sq_plus_one, plain_one) #Encrypted multiplication results in the output ciphertext growing in size. #More precisely, if the input ciphertexts have size M and N, then the output #ciphertext after homomorphic multiplication will have size M+N-1. In this #case we perform a squaring, and observe both size growth and noise budget #consumption. print(" + size of x_sq_plus_one: {}".format(x_sq_plus_one.size())) print(" + noise budget in x_sq_plus_one: {} bits".format( decryptor.invariant_noise_budget(x_sq_plus_one))) #Even though the size has grown, decryption works as usual as long as noise #budget has not reached 0. decrypted_result = Plaintext() decryptor.decrypt(x_sq_plus_one, decrypted_result) print(" + decryption of x_sq_plus_one: 0x{} ...... Correct.".format( decrypted_result.to_string())) #Next, we compute (x + 1)^2. print("Compute x_plus_one_sq ((x+1)^2).") x_plus_one_sq = Ciphertext() evaluator.add_plain(x_encrypted, plain_one, x_plus_one_sq) evaluator.square_inplace(x_plus_one_sq) print(" + size of x_plus_one_sq: {}".format(x_plus_one_sq.size())) print(" + noise budget in x_plus_one_sq: {} bits".format( decryptor.invariant_noise_budget(x_plus_one_sq))) decryptor.decrypt(x_plus_one_sq, decrypted_result) print(" + decryption of x_plus_one_sq: 0x{} ...... Correct.".format( decrypted_result.to_string())) #Finally, we multiply (x^2 + 1) * (x + 1)^2 * 4. print("Compute encrypted_result (4(x^2+1)(x+1)^2).") encrypted_result = Ciphertext() plain_four = Plaintext("4") evaluator.multiply_plain_inplace(x_sq_plus_one, plain_four) evaluator.multiply(x_sq_plus_one, x_plus_one_sq, encrypted_result) print(" + size of encrypted_result: {}".format(encrypted_result.size())) print(" + noise budget in encrypted_result: {} bits".format( decryptor.invariant_noise_budget(encrypted_result))) print("NOTE: Decryption can be incorrect if noise budget is zero.") print("~~~~~~ A better way to calculate 4(x^2+1)(x+1)^2. ~~~~~~") #Noise budget has reached 0, which means that decryption cannot be expected #to give the correct result. This is because both ciphertexts x_sq_plus_one #and x_plus_one_sq consist of 3 polynomials due to the previous squaring #operations, and homomorphic operations on large ciphertexts consume much more #noise budget than computations on small ciphertexts. Computing on smaller #ciphertexts is also computationally significantly cheaper. #`Relinearization' is an operation that reduces the size of a ciphertext after #multiplication back to the initial size, 2. Thus, relinearizing one or both #input ciphertexts before the next multiplication can have a huge positive #impact on both noise growth and performance, even though relinearization has #a significant computational cost itself. It is only possible to relinearize #size 3 ciphertexts down to size 2, so often the user would want to relinearize #after each multiplication to keep the ciphertext sizes at 2. #Relinearization requires special `relinearization keys', which can be thought #of as a kind of public key. Relinearization keys can easily be created with #the KeyGenerator. #Relinearization is used similarly in both the BFV and the CKKS schemes, but #in this example we continue using BFV. We repeat our computation from before, #but this time relinearize after every multiplication. #We use KeyGenerator::relin_keys() to create relinearization keys. print("Generate relinearization keys.") relin_keys = keygen.relin_keys() #We now repeat the computation relinearizing after each multiplication. print("Compute and relinearize x_squared (x^2),") print("then compute x_sq_plus_one (x^2+1)") x_squared = Ciphertext() evaluator.square(x_encrypted, x_squared) print(" + size of x_squared: {}".format(x_squared.size())) evaluator.relinearize_inplace(x_squared, relin_keys) print(" + size of x_squared (after relinearization): {}".format( x_squared.size())) evaluator.add_plain(x_squared, plain_one, x_sq_plus_one) print(" + noise budget in x_sq_plus_one: {} bits".format( decryptor.invariant_noise_budget(x_sq_plus_one))) decryptor.decrypt(x_sq_plus_one, decrypted_result) print(" + decryption of x_sq_plus_one: 0x{} ...... Correct.".format( decrypted_result.to_string())) x_plus_one = Ciphertext() print("Compute x_plus_one (x+1),") print("then compute and relinearize x_plus_one_sq ((x+1)^2).") evaluator.add_plain(x_encrypted, plain_one, x_plus_one) evaluator.square(x_plus_one, x_plus_one_sq) print(" + size of x_plus_one_sq: {}".format(x_plus_one_sq.size())) evaluator.relinearize_inplace(x_plus_one_sq, relin_keys) print(" + noise budget in x_plus_one_sq: {} bits".format( decryptor.invariant_noise_budget(x_plus_one_sq))) decryptor.decrypt(x_plus_one_sq, decrypted_result) print(" + decryption of x_plus_one_sq: 0x{} ...... Correct.".format( decrypted_result.to_string())) print("Compute and relinearize encrypted_result (4(x^2+1)(x+1)^2).") evaluator.multiply_plain_inplace(x_sq_plus_one, plain_four) evaluator.multiply(x_sq_plus_one, x_plus_one_sq, encrypted_result) print(" + size of encrypted_result: {}".format(encrypted_result.size())) evaluator.relinearize_inplace(encrypted_result, relin_keys) print(" + size of encrypted_result (after relinearization): {}".format( encrypted_result.size())) print(" + noise budget in encrypted_result: {} bits".format( decryptor.invariant_noise_budget(encrypted_result))) print("NOTE: Notice the increase in remaining noise budget.") #Relinearization clearly improved our noise consumption. We have still plenty #of noise budget left, so we can expect the correct answer when decrypting. print("Decrypt encrypted_result (4(x^2+1)(x+1)^2).") decryptor.decrypt(encrypted_result, decrypted_result) print(" + decryption of 4(x^2+1)(x+1)^2 = 0x{} ...... Correct.".format( decrypted_result.to_string()))
def example_ckks_basics(): print("Example: CKKS Basics"); #In this example we demonstrate evaluating a polynomial function # # PI*x^3 + 0.4*x + 1 # #on encrypted floating-point input data x for a set of 4096 equidistant points #in the interval [0, 1]. This example demonstrates many of the main features #of the CKKS scheme, but also the challenges in using it. # # We start by setting up the CKKS scheme. parms = EncryptionParameters(scheme_type.CKKS) #We saw in `2_encoders.cpp' that multiplication in CKKS causes scales #in ciphertexts to grow. The scale of any ciphertext must not get too close #to the total size of coeff_modulus, or else the ciphertext simply runs out of #room to store the scaled-up plaintext. The CKKS scheme provides a `rescale' #functionality that can reduce the scale, and stabilize the scale expansion. # #Rescaling is a kind of modulus switch operation (recall `3_levels.cpp'). #As modulus switching, it removes the last of the primes from coeff_modulus, #but as a side-effect it scales down the ciphertext by the removed prime. #Usually we want to have perfect control over how the scales are changed, #which is why for the CKKS scheme it is more common to use carefully selected #primes for the coeff_modulus. # #More precisely, suppose that the scale in a CKKS ciphertext is S, and the #last prime in the current coeff_modulus (for the ciphertext) is P. Rescaling #to the next level changes the scale to S/P, and removes the prime P from the #coeff_modulus, as usual in modulus switching. The number of primes limits #how many rescalings can be done, and thus limits the multiplicative depth of #the computation. # #It is possible to choose the initial scale freely. One good strategy can be #to is to set the initial scale S and primes P_i in the coeff_modulus to be #very close to each other. If ciphertexts have scale S before multiplication, #they have scale S^2 after multiplication, and S^2/P_i after rescaling. If all #P_i are close to S, then S^2/P_i is close to S again. This way we stabilize the #scales to be close to S throughout the computation. Generally, for a circuit #of depth D, we need to rescale D times, i.e., we need to be able to remove D #primes from the coefficient modulus. Once we have only one prime left in the #coeff_modulus, the remaining prime must be larger than S by a few bits to #preserve the pre-decimal-point value of the plaintext. # #Therefore, a generally good strategy is to choose parameters for the CKKS #scheme as follows: # # (1) Choose a 60-bit prime as the first prime in coeff_modulus. This will # give the highest precision when decrypting; # (2) Choose another 60-bit prime as the last element of coeff_modulus, as # this will be used as the special prime and should be as large as the # largest of the other primes; # (3) Choose the intermediate primes to be close to each other. # #We use CoeffModulus::Create to generate primes of the appropriate size. Note #that our coeff_modulus is 200 bits total, which is below the bound for our #poly_modulus_degree: CoeffModulus::MaxBitCount(8192) returns 218. poly_modulus_degree = 8192 parms.set_poly_modulus_degree(poly_modulus_degree) parms.set_coeff_modulus(CoeffModulus.Create( poly_modulus_degree, IntVector([60, 40, 40, 60]))) #We choose the initial scale to be 2^40. At the last level, this leaves us #60-40=20 bits of precision before the decimal point, and enough (roughly #10-20 bits) of precision after the decimal point. Since our intermediate #primes are 40 bits (in fact, they are very close to 2^40), we can achieve #scale stabilization as described above. scale = 2.0**40 context = SEALContext.Create(parms) print_parameters(context) keygen = KeyGenerator(context) public_key = keygen.public_key() secret_key = keygen.secret_key() relin_keys = keygen.relin_keys() encryptor = Encryptor(context, public_key) evaluator = Evaluator(context) decryptor = Decryptor(context, secret_key) encoder = CKKSEncoder(context) slot_count = encoder.slot_count() print("Number of slots: {}".format(slot_count)) step_size = 1.0 / (slot_count - 1) input = DoubleVector(list(map(lambda x: x*step_size, range(0, slot_count)))) print("Input vector: ") print_vector(input) print("Evaluating polynomial PI*x^3 + 0.4x + 1 ...") #We create plaintexts for PI, 0.4, and 1 using an overload of CKKSEncoder::encode #that encodes the given floating-point value to every slot in the vector. plain_coeff3 = Plaintext() plain_coeff1 = Plaintext() plain_coeff0 = Plaintext() encoder.encode(3.14159265, scale, plain_coeff3) encoder.encode(0.4, scale, plain_coeff1) encoder.encode(1.0, scale, plain_coeff0) x_plain = Plaintext() print("Encode input vectors.") encoder.encode(input, scale, x_plain) x1_encrypted = Ciphertext() encryptor.encrypt(x_plain, x1_encrypted) #To compute x^3 we first compute x^2 and relinearize. However, the scale has #now grown to 2^80. x3_encrypted = Ciphertext() print("Compute x^2 and relinearize:") evaluator.square(x1_encrypted, x3_encrypted) evaluator.relinearize_inplace(x3_encrypted, relin_keys) print(" + Scale of x^2 before rescale: {} bits".format(log2(x3_encrypted.scale()))) #Now rescale; in addition to a modulus switch, the scale is reduced down by #a factor equal to the prime that was switched away (40-bit prime). Hence, the #new scale should be close to 2^40. Note, however, that the scale is not equal #to 2^40: this is because the 40-bit prime is only close to 2^40. print("Rescale x^2.") evaluator.rescale_to_next_inplace(x3_encrypted) print(" + Scale of x^2 after rescale: {} bits".format(log2(x3_encrypted.scale()))) #Now x3_encrypted is at a different level than x1_encrypted, which prevents us #from multiplying them to compute x^3. We could simply switch x1_encrypted to #the next parameters in the modulus switching chain. However, since we still #need to multiply the x^3 term with PI (plain_coeff3), we instead compute PI*x #first and multiply that with x^2 to obtain PI*x^3. To this end, we compute #PI*x and rescale it back from scale 2^80 to something close to 2^40. print("Compute and rescale PI*x.") x1_encrypted_coeff3 = Ciphertext() evaluator.multiply_plain(x1_encrypted, plain_coeff3, x1_encrypted_coeff3) print(" + Scale of PI*x before rescale: {} bits".format(log2(x1_encrypted_coeff3.scale()))) evaluator.rescale_to_next_inplace(x1_encrypted_coeff3) print(" + Scale of PI*x after rescale: {} bits".format(log2(x1_encrypted_coeff3.scale()))) #Since x3_encrypted and x1_encrypted_coeff3 have the same exact scale and use #the same encryption parameters, we can multiply them together. We write the #result to x3_encrypted, relinearize, and rescale. Note that again the scale #is something close to 2^40, but not exactly 2^40 due to yet another scaling #by a prime. We are down to the last level in the modulus switching chain. print("Compute, relinearize, and rescale (PI*x)*x^2.") evaluator.multiply_inplace(x3_encrypted, x1_encrypted_coeff3) evaluator.relinearize_inplace(x3_encrypted, relin_keys) print(" + Scale of PI*x^3 before rescale: {} bits".format(log2(x3_encrypted.scale()))) evaluator.rescale_to_next_inplace(x3_encrypted) print(" + Scale of PI*x^3 after rescale: {} bits".format(log2(x3_encrypted.scale()))) #Next we compute the degree one term. All this requires is one multiply_plain #with plain_coeff1. We overwrite x1_encrypted with the result. print("Compute and rescale 0.4*x.") evaluator.multiply_plain_inplace(x1_encrypted, plain_coeff1) print(" + Scale of 0.4*x before rescale: {} bits".format(log2(x1_encrypted.scale()))) evaluator.rescale_to_next_inplace(x1_encrypted) print(" + Scale of 0.4*x after rescale: {} bits".format(log2(x1_encrypted.scale()))) #Now we would hope to compute the sum of all three terms. However, there is #a serious problem: the encryption parameters used by all three terms are #different due to modulus switching from rescaling. # #Encrypted addition and subtraction require that the scales of the inputs are #the same, and also that the encryption parameters (parms_id) match. If there #is a mismatch, Evaluator will throw an exception. print("Parameters used by all three terms are different.") print(" + Modulus chain index for x3_encrypted: {}".format( context.get_context_data(x3_encrypted.parms_id()).chain_index())) print(" + Modulus chain index for x1_encrypted: {}".format( context.get_context_data(x1_encrypted.parms_id()).chain_index())) print(" + Modulus chain index for plain_coeff0: {}".format( context.get_context_data(plain_coeff0.parms_id()).chain_index())) #Let us carefully consider what the scales are at this point. We denote the #primes in coeff_modulus as P_0, P_1, P_2, P_3, in this order. P_3 is used as #the special modulus and is not involved in rescalings. After the computations #above the scales in ciphertexts are: # # - Product x^2 has scale 2^80 and is at level 2; # - Product PI*x has scale 2^80 and is at level 2; # - We rescaled both down to scale 2^80/P_2 and level 1; # - Product PI*x^3 has scale (2^80/P_2)^2; # - We rescaled it down to scale (2^80/P_2)^2/P_1 and level 0; # - Product 0.4*x has scale 2^80; # - We rescaled it down to scale 2^80/P_2 and level 1; # - The contant term 1 has scale 2^40 and is at level 2. # #Although the scales of all three terms are approximately 2^40, their exact #values are different, hence they cannot be added together. print("The exact scales of all three terms are different:") print(" + Exact scale in PI*x^3: {0:0.10f}".format(x3_encrypted.scale())) print(" + Exact scale in 0.4*x: {0:0.10f}".format(x1_encrypted.scale())) print(" + Exact scale in 1: {0:0.10f}".format(plain_coeff0.scale())) #There are many ways to fix this problem. Since P_2 and P_1 are really close #to 2^40, we can simply "lie" to Microsoft SEAL and set the scales to be the #same. For example, changing the scale of PI*x^3 to 2^40 simply means that we #scale the value of PI*x^3 by 2^120/(P_2^2*P_1), which is very close to 1. #This should not result in any noticeable error. # #Another option would be to encode 1 with scale 2^80/P_2, do a multiply_plain #with 0.4*x, and finally rescale. In this case we would need to additionally #make sure to encode 1 with appropriate encryption parameters (parms_id). # #In this example we will use the first (simplest) approach and simply change #the scale of PI*x^3 and 0.4*x to 2^40. print("Normalize scales to 2^40.") x3_encrypted.set_scale(2.0**40) x1_encrypted.set_scale(2.0**40) #We still have a problem with mismatching encryption parameters. This is easy #to fix by using traditional modulus switching (no rescaling). CKKS supports #modulus switching just like the BFV scheme, allowing us to switch away parts #of the coefficient modulus when it is simply not needed. print("Normalize encryption parameters to the lowest level.") last_parms_id = x3_encrypted.parms_id() evaluator.mod_switch_to_inplace(x1_encrypted, last_parms_id) evaluator.mod_switch_to_inplace(plain_coeff0, last_parms_id) #All three ciphertexts are now compatible and can be added. print("Compute PI*x^3 + 0.4*x + 1.") encrypted_result = Ciphertext() evaluator.add(x3_encrypted, x1_encrypted, encrypted_result) evaluator.add_plain_inplace(encrypted_result, plain_coeff0) #First print the true result. plain_result = Plaintext() print("Decrypt and decode PI*x^3 + 0.4x + 1.") print(" + Expected result:") true_result = DoubleVector(list(map(lambda x: (3.14159265 * x * x + 0.4)* x + 1, input))) print_vector(true_result) #Decrypt, decode, and print the result. decryptor.decrypt(encrypted_result, plain_result) result = DoubleVector() encoder.decode(plain_result, result) print(" + Computed result ...... Correct.") print_vector(result)
class SealOps: @classmethod def with_env(cls): parms = EncryptionParameters(scheme_type.CKKS) parms.set_poly_modulus_degree(POLY_MODULUS_DEGREE) parms.set_coeff_modulus( CoeffModulus.Create(POLY_MODULUS_DEGREE, PRIME_SIZE_LIST)) context = SEALContext.Create(parms) keygen = KeyGenerator(context) public_key = keygen.public_key() secret_key = keygen.secret_key() relin_keys = keygen.relin_keys() galois_keys = keygen.galois_keys() return cls(context=context, public_key=public_key, secret_key=secret_key, relin_keys=relin_keys, galois_keys=galois_keys, poly_modulus_degree=POLY_MODULUS_DEGREE, scale=SCALE) def __init__(self, context: SEALContext, scale: float, poly_modulus_degree: int, public_key: PublicKey = None, secret_key: SecretKey = None, relin_keys: RelinKeys = None, galois_keys: GaloisKeys = None): self.scale = scale self.context = context self.encoder = CKKSEncoder(context) self.evaluator = Evaluator(context) self.encryptor = Encryptor(context, public_key) self.decryptor = Decryptor(context, secret_key) self.relin_keys = relin_keys self.galois_keys = galois_keys self.poly_modulus_degree_log = np.log2(poly_modulus_degree) def encrypt(self, matrix: np.array): matrix = Matrix.from_numpy_array(array=matrix) cipher_matrix = CipherMatrix(rows=matrix.rows, cols=matrix.cols) for i in range(matrix.rows): encoded_row = Plaintext() self.encoder.encode(matrix[i], self.scale, encoded_row) self.encryptor.encrypt(encoded_row, cipher_matrix[i]) return cipher_matrix def decrypt(self, cipher_matrix: CipherMatrix) -> Matrix: matrix = Matrix(rows=cipher_matrix.rows, cols=cipher_matrix.cols) for i in range(matrix.rows): row = Vector() encoded_row = Plaintext() self.decryptor.decrypt(cipher_matrix[i], encoded_row) self.encoder.decode(encoded_row, row) matrix[i] = row return matrix def add(self, matrix_a: CipherMatrix, matrix_b: CipherMatrix) -> CipherMatrix: self.validate_same_dimension(matrix_a, matrix_b) result_matrix = CipherMatrix(rows=matrix_a.rows, cols=matrix_a.cols) for i in range(matrix_a.rows): a_tag, b_tag = self.get_matched_scale_vectors( matrix_a[i], matrix_b[i]) self.evaluator.add(a_tag, b_tag, result_matrix[i]) return result_matrix def add_plain(self, matrix_a: CipherMatrix, matrix_b: np.array) -> CipherMatrix: matrix_b = Matrix.from_numpy_array(matrix_b) self.validate_same_dimension(matrix_a, matrix_b) result_matrix = CipherMatrix(rows=matrix_a.rows, cols=matrix_a.cols) for i in range(matrix_a.rows): row = matrix_b[i] encoded_row = Plaintext() self.encoder.encode(row, self.scale, encoded_row) self.evaluator.mod_switch_to_inplace(encoded_row, matrix_a[i].parms_id()) self.evaluator.add_plain(matrix_a[i], encoded_row, result_matrix[i]) return result_matrix def multiply_plain(self, matrix_a: CipherMatrix, matrix_b: np.array) -> CipherMatrix: matrix_b = Matrix.from_numpy_array(matrix_b) self.validate_same_dimension(matrix_a, matrix_b) result_matrix = CipherMatrix(rows=matrix_a.rows, cols=matrix_a.cols) for i in range(matrix_a.rows): row = matrix_b[i] encoded_row = Plaintext() self.encoder.encode(row, self.scale, encoded_row) self.evaluator.mod_switch_to_inplace(encoded_row, matrix_a[i].parms_id()) self.evaluator.multiply_plain(matrix_a[i], encoded_row, result_matrix[i]) return result_matrix def dot_vector(self, a: Ciphertext, b: Ciphertext) -> Ciphertext: result = Ciphertext() self.evaluator.multiply(a, b, result) self.evaluator.relinearize_inplace(result, self.relin_keys) self.vector_sum_inplace(result) self.get_vector_first_element(result) self.evaluator.rescale_to_next_inplace(result) return result def dot_vector_with_plain(self, a: Ciphertext, b: DoubleVector) -> Ciphertext: result = Ciphertext() b_plain = Plaintext() self.encoder.encode(b, self.scale, b_plain) self.evaluator.multiply_plain(a, b_plain, result) self.vector_sum_inplace(result) self.get_vector_first_element(result) self.evaluator.rescale_to_next_inplace(result) return result def get_vector_range(self, vector_a: Ciphertext, i: int, j: int) -> Ciphertext: cipher_range = Ciphertext() one_and_zeros = DoubleVector([0.0 if x < i else 1.0 for x in range(j)]) plain = Plaintext() self.encoder.encode(one_and_zeros, self.scale, plain) self.evaluator.mod_switch_to_inplace(plain, vector_a.parms_id()) self.evaluator.multiply_plain(vector_a, plain, cipher_range) return cipher_range def dot_matrix_with_matrix_transpose(self, matrix_a: CipherMatrix, matrix_b: CipherMatrix): result_matrix = CipherMatrix(rows=matrix_a.rows, cols=matrix_a.cols) rows_a = matrix_a.rows cols_b = matrix_b.rows for i in range(rows_a): vector_dot_products = [] zeros = Plaintext() for j in range(cols_b): vector_dot_products += [ self.dot_vector(matrix_a[i], matrix_b[j]) ] if j == 0: zero = DoubleVector() self.encoder.encode(zero, vector_dot_products[j].scale(), zeros) self.evaluator.mod_switch_to_inplace( zeros, vector_dot_products[j].parms_id()) self.evaluator.add_plain(vector_dot_products[j], zeros, result_matrix[i]) else: self.evaluator.rotate_vector_inplace( vector_dot_products[j], -j, self.galois_keys) self.evaluator.add_inplace(result_matrix[i], vector_dot_products[j]) for vec in result_matrix: self.evaluator.relinearize_inplace(vec, self.relin_keys) self.evaluator.rescale_to_next_inplace(vec) return result_matrix def dot_matrix_with_plain_matrix_transpose(self, matrix_a: CipherMatrix, matrix_b: np.array): matrix_b = Matrix.from_numpy_array(matrix_b) result_matrix = CipherMatrix(rows=matrix_a.rows, cols=matrix_a.cols) rows_a = matrix_a.rows cols_b = matrix_b.rows for i in range(rows_a): vector_dot_products = [] zeros = Plaintext() for j in range(cols_b): vector_dot_products += [ self.dot_vector_with_plain(matrix_a[i], matrix_b[j]) ] if j == 0: zero = DoubleVector() self.encoder.encode(zero, vector_dot_products[j].scale(), zeros) self.evaluator.mod_switch_to_inplace( zeros, vector_dot_products[j].parms_id()) self.evaluator.add_plain(vector_dot_products[j], zeros, result_matrix[i]) else: self.evaluator.rotate_vector_inplace( vector_dot_products[j], -j, self.galois_keys) self.evaluator.add_inplace(result_matrix[i], vector_dot_products[j]) for vec in result_matrix: self.evaluator.relinearize_inplace(vec, self.relin_keys) self.evaluator.rescale_to_next_inplace(vec) return result_matrix @staticmethod def validate_same_dimension(matrix_a, matrix_b): if matrix_a.rows != matrix_b.rows or matrix_a.cols != matrix_b.cols: raise ArithmeticError("Matrices aren't of the same dimension") def vector_sum_inplace(self, cipher: Ciphertext): rotated = Ciphertext() for i in range(int(self.poly_modulus_degree_log - 1)): self.evaluator.rotate_vector(cipher, pow(2, i), self.galois_keys, rotated) self.evaluator.add_inplace(cipher, rotated) def get_vector_first_element(self, cipher: Ciphertext): one_and_zeros = DoubleVector([1.0]) plain = Plaintext() self.encoder.encode(one_and_zeros, self.scale, plain) self.evaluator.multiply_plain_inplace(cipher, plain) def get_matched_scale_vectors(self, a: Ciphertext, b: Ciphertext) -> (Ciphertext, Ciphertext): a_tag = Ciphertext(a) b_tag = Ciphertext(b) a_index = self.context.get_context_data(a.parms_id()).chain_index() b_index = self.context.get_context_data(b.parms_id()).chain_index() # Changing the mod if required, else just setting the scale if a_index < b_index: self.evaluator.mod_switch_to_inplace(b_tag, a.parms_id()) elif a_index > b_index: self.evaluator.mod_switch_to_inplace(a_tag, b.parms_id()) a_tag.set_scale(self.scale) b_tag.set_scale(self.scale) return a_tag, b_tag