ANALYSIS OF RSA ALGORITHM USING GPU PROGRAMMING

International Journal of Network Security & Its Applications (IJNSA), Vol.6, No.4, July 2014
DOI : 10.5121/ijnsa.2014.6402 13
ANALYSIS OF RSA ALGORITHM USING GPU
PROGRAMMING
Sonam Mahajan1
and Maninder Singh2
1
Department of Computer Science Engineering, Thapar University, Patiala, India
2
Department of Computer Science Engineering, Thapar University, Patiala, India
ABSTRACT
Modern-day computer security relies heavily on cryptography as a means to protect the data that we have
become increasingly reliant on. The main research in computer security domain is how to enhance the
speed of RSA algorithm. The computing capability of Graphic Processing Unit as a co-processor of the
CPU can leverage massive-parallelism. This paper presents a novel algorithm for calculating modulo
value that can process large power of numbers which otherwise are not supported by built-in data types.
First the traditional algorithm is studied. Secondly, the parallelized RSA algorithm is designed using
CUDA framework. Thirdly, the designed algorithm is realized for small prime numbers and large prime
number . As a result the main fundamental problem of RSA algorithm such as speed and use of poor or
small prime numbers that has led to significant security holes, despite the RSA algorithm's mathematical
soundness can be alleviated by this algorithm.
KEYWORDS
CPU, GPU, CUDA, RSA, Cryptographic Algorithm.
1. INTRODUCTION
RSA (named for its inventors, Ron Rivest, Adi Shamir, and Leonard Adleman [1] ) is a public
key encryption scheme. This algorithm relies on the difficulty of factoring large numbers which
has seriously affected its performance and so restricts its use in wider applications. Therefore, the
rapid realization and parallelism of RSA encryption algorithm has been a prevalent research
focus. With the advent of CUDA technology, it is now possible to perform general-purpose
computation on GPU [2]. The primary goal of our work is to speed up the most computationally
intensive part of their process by implementing the GCD comparisons of RSA keys using
NVIDIA's CUDA platform.
The reminder of this paper is organized as follows. In section 2,3,4, we study the traditional RSA
algorithm. In section 5, we explained our system hardware. In section 6,7,8, 9 we explained the
design and implementation of parallelized algorithm. Section 10 gives the result of our
parallelized algorithm and section 12 concludes the paper.
2. TRADITIONAL RSA ALGORITHM[1]
RSA is an algorithm for public-key cryptography [1] and is considered as one of the great
advances in the field of public key cryptography. It is suitable for both signing and encryption.
Electronic commerce protocols mostly rely on RSA for security. Sufficiently long keys and up-to-
date implementation of RSA is considered more secure to use.

14
RSA is an asymmetric key encryption scheme which makes use of two different keys for
encryption and decryption. The public key that is known to everyone is used for encryption. The
messages encrypted using the public key can only be decrypted by using private key. The key
generation process of RSA algorithm is as follows:
The public key is comprised of a modulus n of specified length (the product of primes p and q),
and an exponent e. The length of n is given in terms of bits, thus the term “8-bit RSA key" refers
to the number of bits which make up this value. The associated private key uses the same n, and
another value d such that d*e = 1 mod φ (n) where φ (n) = (p - 1)*(q - 1) [3]. For a plaintext M
and cipher text C, the encryption and decryption is done as follows:
C= Me
mod n, M = Cd
mod n.
For example, the public key (e, n) is (131,17947), the private key (d, n) is (137,17947), and let
suppose the plaintext M to be sent is: parallel encryption.
• Firstly, the sender will partition the plaintext into packets as: pa ra ll el en cr yp ti on. We
suppose a is 00, b is 01, c is 02, ..... z is 25.
• Then further digitalize the plaintext packets as: 1500 1700 1111 0411 0413 0217 2415
1908 1413.
• After that using the encryption and decryption transformation given above calculate the
cipher text and the plaintext in digitalized form.
• Convert the plaintext into alphabets, which is the original: parallel encryption.
3. OPERATION [1]
The three steps of RSA algorithm – Key Generation, Encryption and Decryption are explained as
follows:
3.1 Key generation
As explained above RSA is an asymmetric key encryption scheme. It makes use of two different
keys for encryption and decryption. The public key that is known to everyone is used for
encryption. The messages encrypted using the public key can only be decrypted by using private
key. The keys for RSA algorithm are generated as follows:
Figure 1. RSA Key Generation

15
3.2 Encryption process
Alice transmits its public key (n,e) to the Bob and keeps the private key secret. Suppose Bob
wishes to send message M to Alice. He first convert the message M to an integer m , such that
m is between 0 and n using padding scheme [1][3]. Finally cipher text c is calculated as: C = me
mod n. This cipher text is then transmitted to Alice.
3.3 Decryption process
Alice can decode the cipher text using her private key component d as follows:
m = Cd mod n
Given m, original message M can be recovered.
3.4 Example
Figure 2. RSA Example
3.6 Modular exponential
3.6.1 Modular arithmetic
Public key cryptography is computationally expensive as it mostly includes raising a large power
to a base and then reducing the result using modulo function. This process is known as modular
exponential. In practical, RSA algorithm should be fast , i.e., modular exponential function should
be fast and along with it should be efficient. The simple modular operations are given in the
Figure 3.
Figure 3. Modular Arithmetic
3.6.2 Naive modular exponentiation
In this method modular multiplications are applied repeatedly. For example: with g=4, e=13 and
m= 497, the naive modular exponential will solve ge mod m as shown in figure and gives the
result as 445. This method is not efficient as it performs e-1 modular multiplications. In
cryptography the security depends on the larger value of e and efficiency depends how efficiently
these modular multiplications and modular exponential functions are solved. The plaintext ,

16
cipher-text or even partial cipher text is supposed to be of large in value and it will require a large
amount of modular multiplications if we rely on this naive algorithm.
Figure 4. Naive Modular Exponentiation Example
3.6.3 Repeated square-and-multiply methods
It makes use of the fact that if the e value is even, then the modular exponential is calculated as
ge mod m =( ge/2 * ge/2) mod m ,hence by reducing the amount of modular multiplications to 2z
where z is the number of bits when converting e to binary form. The algorithm comes with two
forms as shown below.
• Right-to-left binary modular exponential
• Left-to-right binary modular exponential
This algorithm works efficiently for large value of e.

17
Figure 5. Repeated Square-and-Multiply Methods
3.6.4 Left-to-right k-ary exponentiation
This algorithm is the generalization of the previous left-to-right binary modular exponentiation
described previously. In this algorithm each iteration processes more than one bit of the exponent
and works efficiently when pre-computations are done in advance and is used again and again.
Figure 6. Left-To-Right K-ary Modular Exponentiation
3.6.5 Sliding window exponential
The main idea behind this algorithm is the reduction of pre-computations as compared to the
above discussed algorithm and hence ultimately reduction in the average number of
multiplications done.

18
Figure 7. Sliding Window Exponential
3.7 ADVANTAGES AND DISADVANTAGES OF RSA ALGORITHM
The RSA algorithm is known for its increased security and convenience. In RSA algorithm
private keys are neither transmitted or nor revealed to anyone. By contrast in secret-key system
keys are exchanged either manually or with the help of communication channel. This is because
in secret-key system same key is used for encryption and decryption and there exist a chance that
an enemy can retrieve this secret key during transmission
Public-key systems can provide digital signatures. These digital signatures cannot be repudiated.
By contrast in secret-key systems authentication requires either sharing of some important secrets
or the involvement of third party. This can result in sender repudiating the previously
authenticated message by claiming that the shared secret is compromised any one of the party
involved
Speed is the main disadvantage of public-key cryptography. Because there is always a trade off
between efficiency and security. For efficient RSA, key size should be small but small key size
leads to many security holes. There exist many secret-key methods that are faster than the
asymmetric encryption method. In reality public-key encryption is used to supplement the
private-key encryption.
4. GPU-ACCELERATED COMPUTING[4][3]
Cryptographic operations are highly computationally expensive and hence consume a large
amount of CPU cycles from the system. Such examples are web servers. To incorporate with this
and to offload cryptographic operations, web servers are typically deployed in systems using
cryptographic accelerator cards. Though this saves system CPU cycles from complex and
computationally intensive application logic but it also results in difficult deployment scenarios.
So, the possibility of offloading the massive computationally rich operations to GPU would
lighten the CPU load to a great extent and hence the saving can be used for other applications.

19
The idea also provide a significant performance increase. This technique is known as “GPU
ACCELERATED COMPUTING”. The idea will be clear from the Figure. 8
Figure 8. GPU Accelerated Computing
5. ARCHITECTURE OVERVIEW[4]
NVIDIA's Compute Unified Device Architecture (CUDA) platform provides a set of tools to
write programs that make use of NVIDIA's GPUs [3]. These massively-parallel hardware devices
process large amounts of data simultaneously and allow significant speedups in programs with
sections of parallelizable code making use of the Simultaneous Program, Multiple Data (SPMD)
model.
Figure 9. CUDA System Model
The platform allows various arrangements of threads to perform work, according to the
developer's problem decomposition. In general, individual threads are grouped into up-to 3-
dimensional blocks to allow sharing of common memory between threads. These blocks can then

20
further be organized into a 2-dimensional grid. The GPU breaks the total number of threads into
groups called warps, which, on current GPU hardware, consist of 32 threads that will be executed
simultaneously on a single streaming multiprocessor (SM). The GPU consists of several SMs
which are each capable of executing a warp. Blocks are scheduled to SMs until all allocated
threads have been executed. There is also a memory hierarchy on the GPU. Three of the various
types of memory are relevant to this work: global memory is the slowest and largest; shared
memory is much faster, but also significantly smaller; and a limited number of registers that each
SM has access to. Each thread in a block can access the same section of shared memory.
6. PARALLELIZATION
The algorithm used to parallelize the RSA modulo function works as follows:
• CPU accepts the values of the message and the key parameters.
• CPU allocates memory on the CUDA enabled device and copies the values on the device
• CPU invokes the CUDA kernel on the GPU
• GPU encrypts each message character with RSA algorithm with the number of threads
equal to the message length.
• The control is transferred back to CPU
• CPU copies and displays the results from the GPU.
As per the flow given above the kernel is so built to calculate the cipher text C = Me
mod n. The
kernel so designed works efficiently and uses the novel algorithm for calculating modulo value.
The algorithm for calculating modulo value is implemented such that it can hold for very large
power of numbers which are not supported by built in data types. The modulus value is calculated
using the following principle:
• C = Me
mod n
• C = (Me-x
mod n * Mx
mod n) mod n
Hence iterating over a suitable value of x gives the desired result.
7. ALGORITHM FOR CUDA KERNEL
Algorithm 1:

21
Algorithm 2:
In the first algorithm we, first pass the arguments as pointer to global rsa(). After declaration of
the variables we assign the thread index. Thread index is the unique id of the each thread in a
block that will execute the same instruction but with different data sets that will be distributed
among the threads. After that in the 4th
Step, check the condition if the number of threads to be
executed is greater than the threads assigned, call Algorithm 2 for mod() with three parameters.
Finally after the execution of the Algorithm 2, encrypted text will be handled to the host as
result[] array.
In Algorithm 1 __global__ keywords represents that the rsa() will execute on the device and will
be called from the host. We use pointers for the variables because rsa() runs on device, so
variables must point to device memory. For that we need to allocate memory on GPU
This algorithm works for large size of values which are not otherwise supported by the built-in
data type.

22
8. Integration of C with CUDA API
Figure 10. CUDA-C Host Code for RSA Algorithm
The above is the small part of the host code showing the integration of CPU and GPU.
• Check for available GPU devices;
• Initialize host and device copies of variables;
• Allocate space for device copies of variables using cudaMalloc();
• Setup input values;
• Copy input values to device using cudaMemcpy;
• Launch kernel on GPU;
rsa<<<nblocks, nthreads>>>(dev_num,dev_key,dev_den,dev_res)
• Copy result back to host using cudaMemcpy();
• Cleanup.
In the first algorithm we, first pass the arguments as pointer to global rsa(). After declaration of
the variables we assign the thread index. Thread index is the unique id of the each thread in a
block that will execute the same instruction but with different data sets that will be distributed
among the threads. After that in the 4th
Step, check the condition if the number of threads to be
executed is greater than the threads assigned, call Algorithm 2 for mod() with three parameters.
Finally after the execution of the Algorithm 2, encrypted text will be handled to the host as
result[] array.
In Algorithm 1 __global__ keywords represents that the rsa() will execute on the device and will
be called from the host. We use pointers for the variables because rsa() runs on device, so
variables must point to device memory. For that we need to allocate memory on GPU
This algorithm works for large size of values which are not otherwise supported by the built-in
data type.
9. KERNEL CODE
As introduced in section 2, RSA algorithm divides the plaintext or cipher text into packets of
same length and then apply encryption or decryption transformation on each packet. A question is
how does a thread know which elements are assigned to it and are supposed to process them?
CUDA user can get the thread and block index of the thread call it in the function running on

23
device. In this level, the CUDA Multi-threaded programming model will dramatically enhanced
the speed of RSA algorithm. The experimental results will be showed in section 10.
The kernel code used in our experiment is shown below. First CUDA user assign the thread and
block index, so as to let each thread know which elements they are supposed to process. It is
shown in Figure 11. Then it call for another device function to calculate the most intense part of
the RSA algorithm. Note in the below figure11 and figure12, it works for 3 elements.
Figure 11. Kernel code
Figure 12. Kernel’s Device code
10. VERIFICATION
In this section we setup the test environment and design three tests. At first test, we develop a
program running in traditional mode for small prime numbers (only use CPU for computing).
And at the second test, we use CUDA framework to run the RSA algorithm for small prime
numbers in multiple-thread mode. Comparison is done between the two test cases and speed up is
calculated. In the third test we run the GPU RSA for large prime numbers that is not supported
by the built-in data types of CPU RSA. The test result will be showed in this section

24
10.1. Test environment
The code has been tested for :
• Values of message between 0 and 800 which can accommodate the complete ASCII
table
• 8 bit Key Values
The computer we used for testing has an Intel(R) Core(TM) i3-2370M 2.4GHZ CPU, 4 GB
RAM, Windows 7OS and a Nvidia GeForce GT 630M with 512MB memory, and a 2GHZ DDR2
memory. At the first stage, we use Visual Studio 2010 for developing and testing the traditional
RSA algorithm using C language for small prime numbers. Series of input data used for testing
and the result will be showed later.
At the second stage, we also use Visual Studio 2010 for developing and testing parallelized RSA
developed using CUDA v5.5 for small prime numbers. After that the results of stage one and
stage second are compared and hence calculating the respective speedup.
In the third test we run the GPU RSA for large prime numbers that is not supported by the built-in
data types of CPU RSA. The test result will be showed in this section. At present the calculation
of Cipher text using an 8-bit key has been implemented parallel on an array of integers.
10.2 Results
In this part, we show the experiment results for GPU RSA and CPU RSA for small value of n.
Table 1. Comparison of CPU RSA and GPU RSA for small prime numbers i.e (n=131*137)
Data
Size(bytes)
No. of blocks Threads per
block
GPU RSA
Time
CPU RSA
Time
Speedup
256 4 64 7.56 12.56 1.66
512 8 64 7.25 19.14 2.65
1024 16 64 6.86 23.60 3.44
2048 32 64 5.38 29.33 5.51
4096 64 64 5.68 32.64 5.74
8192 128 64 6.27 35.16 5.60
16392 256 64 7.21 39.66 5.50
32784 512 64 9.25 42.37 4.58
Table 1 shows the relationship between the amount of data inputting to the RSA algorithm and
the execution times (in seconds) in traditional mode and multiple thread mode. The first column
shows the number of the data input to the algorithm, and the second column shows the number of
blocks used to process the data input. In the above table 64 threads per block are used to execute
RSA. The execution time is calculated in seconds. In the last column speed up is calculated.
Above results are calculated by making average of the results so taken 20 times to have final
more accurate and precise results.
The enhancement of the execution performance using CUDA framework can be visually
demonstrated by Figure 13.

25
Figure 13. Graph showing effect of data input on CPU RSA and GPU RSA along with the Speedup
10.2.1 GPU RSA for large prime numbers
In this part, we show the experiment results for GPU RSA and CPU RSA for small value of n.
Table 2. GPU RSA for large prime numbers and large value of n (n = 1005 * 509)
Data Size(bytes) No. of blocks Threads per block GPU RSA Time
256 8 32 6.08
512 16 32 6.52
1024 32 32 6.69
2048 64 32 5.53
4096 128 32 6.58
8192 256 32 6.66
16392 512 32 7.81
32784 1024 32 8.76
From Table 2, we can see the relationship between the execution time in seconds and the input
data amount (data in bytes) is linear for certain amount of input. When we use 256 data size to
execute the RSA algorithm, the execution time is very short as compared to traditional mode
which is clearly proved in the above section where the comparison is made for CPU RSA and
GPU RSA for small prime numbers and hence for small value of n. So we can say when the data
size increases, the running time will be significantly reduced depending upon the number of
threads used. Furthermore, we also find that when the data size increases from 1024 to 8192, the
execution time of 7168 threads almost no increase, which just proves our point of view, the more
the data size is, the higher the parallelism of the algorithm, and the shorter the time spent.
Execution time varies according the number of threads and number of blocks used for data input.
In the above table threads per block are constant i.e we use 32 threads per block and number of
blocks used are adjusted according to the data input.
The enhancement of the execution performance of data input in bytes using the large value of
prime numbers (n=1009 * 509) and hence large value of n on CUDA framework can be visually
demonstrated by Figure 14.

26
Figure 14. GPU RSA for large value of n (n=1009*509)
10.2.2. Execution time comparison of GPU RSA for large value of n (1009*509) with
CPU RSA for small value of n(137*131)
In the third and final stage of test results analysis, we analyse our results between
sequential RSA that is using small value of n (17947) and parallelized RSA that is
making use of large prime numbers and large value of n (513581). The enhancement of
the GPU execution performance of data input in bytes using the large value of prime
numbers (n=1009 * 509) on CUDA framework and CPU RSA using small value of prime
numbers (n=137*131) can be visually demonstrated by Figure 15. Hence, we can
leverage massive-parallelism and the computational power that is granted by today's
commodity hardware such as GPUs to make checks that would otherwise be impossible
to perform, attainable.
Figure 15. Comparison of CPU RSA for small prime numbers with GPU RSA for large prime numbers.

27
11. RSA DECRYPTION USING CUDA-C
In this paper, we presented our experience of porting RSA encryption algorithm on to CUDA
architecture. We analyzed the parallel RSA encryption algorithm. As explained above the
encryption and decryption process is done as follows:
C= Me
mod n, M = Cd
mod n.
The approach used for encryption process is same for decryption too. Same kernel code will work
for decryption too. The only parameters that will change is the private key (d) and ciphertext in
place of message bits used during encryption.
12. CONCLUSIONS
In this paper, we presented our experience of porting RSA algorithm on to CUDA architecture.
We analyzed the parallel RSA algorithm. The bottleneck for RSA algorithm lies in the data size
and key size i.e the use of large prime numbers. The use of small prime numbers make RSA
vulnerable and the use of large prime numbers for calculating n makes it slower as computation
expense increases. This paper design a method to computer the data bits parallel using the threads
respectively based on CUDA. This is in order to realize performance improvements which lead to
optimized results.
In the next work, we encourage ourselves to focus on implementation of GPU RSA for large key
size including modular exponentiation algorithms. As it will drastically increase the security in
the public-key cryptography. GPU are becoming popular so deploying cryptography on new
platforms will be very useful.
REFERENCES
[1] R. L. Rivest, A. Shamir, and L. Adleman. A method for obtaining digital signatures and public-
key cryptosystems. Communications of the ACM, 21(2):120{126, 1978.
[2] J. Owens, D. Luebke, N. Govindaraju.. A survey of general-purpose computation on graphics
hardware. Computer Graphics Forum, 26(1): 80{113 , March 2007.
[3] N. Heninger, Z. Durumeric, E. Wustrow, and J. A. Halderman. Mining your Ps and Qs: detection
of widespread weak keys in network devices. In Proceedings of the 21st USENIX conference on
Security symposium, pages 205{220. USENIX Association, 2012 .
[4] NVIDIA|CUDA documents |Programming Guide |CUDA v5.5.
http://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf

28
Authors
Sonam Mahajan
Student , ME – Information Security
Computer Science and Engineering Department
Thapar University
Patiala-147004
Dr. Maninder Singh
Associate Professor
Computer Science and Engineering Department
Thapar University
Patiala-147004

ANALYSIS OF RSA ALGORITHM USING GPU PROGRAMMING

Recommended

More Related Content

Similar to ANALYSIS OF RSA ALGORITHM USING GPU PROGRAMMING (20)

Recently uploaded (20)

ANALYSIS OF RSA ALGORITHM USING GPU PROGRAMMING