Equipped with the latest image processing systems, sensors, and high-resolution cameras, they can conduct real-time aerial photography, monitor enemy activity, and gather critical intelligence without putting the military at risk. UAVs make it possible to conduct long-term operations in conditions of secrecy, providing commanders with valuable information for making strategic decisions. However, the issue of ensuring the confidentiality of critical data (images) collected using UAVs remains unresolved. In this paper universal dataset of cryptographic algorithms is proposed and it uses a neural network model to select the optimal encryption algorithm. To form such a dataset, it was necessary to evaluate the speed and security of the cryptographic algorithms as well as other important parameters. The developed dataset in synthesis with a neural network model can be used to select the optimal crypto algorithm. In further research, the authors plan to determine the criteria for using the generated dataset by neural networks and develop a knowledge base for neural network training.
In the modern world, unmanned aerial vehicles
(UAV) play a key role in many areas of human
activity – from civil services and logistics to
reconnaissance and warfare [
To provide a high level of data confidentiality cryptographic algorithms can be used. There are many advanced directions and technologies in cryptography such as post-quantum cryptography, quantum key distribution and quantum secure direct communication, lightweight cryptography, etc.
Many of these cryptographic methods and protocols can be used in UAV-based systems for image encryption. However, its usage depends on many factors and parameters that should be analyzed for the development of the universal cryptosystem.
In [
According to this, the work aims to ensure the
confidentiality of data during transmission from
UAVs due to the use of neural networks and the
creation of a universal dataset of modern
cryptographic algorithms [
For the effective formation of the dataset, it is also necessary to evaluate the speed of cryptoalgorithms, their crypto-resistance, etc. In the future, the developed dataset can be used in combination with a neural network to select the optimal encryption algorithm depending on the operating conditions of the UAV.
To ensure the protection of information in modern information and communication systems, a large number of crypto algorithms are presented, which differ in their basic characteristics (parameters) - cryptoresistance (to various known methods of cryptanalysis), speed of cryptographic data processing, convenience of software/hardware implementation, etc. Given the hardware limitations of modern UAVs, the speed factor and the amount of resources needed to encrypt information become key characteristics when choosing an algorithm. Among the algorithms that, according to the authors, should be represented in the dataset, the following symmetric stream and block ciphers were chosen: Salsa20; PANAMA; HC-256; AES; DES; Triple DES; Serpent; Blowfish; Twofish; MARS; RC2; RC6; GOST 28147-89 (DSTU GOST 28147: 2009); Kalyna. 3.1.1 Salsa20
A crypto algorithm designed by D. Bernstein and presented at the eSTREAM competition, the purpose of which was to create European standards for data encryption transmitted in postal systems. The algorithm became the winner of the competition in the first category (stream ciphers for high-bandwidth software applications). For implementation, it is necessary to create a key with a length of 128 or 256 bits, as well as an initialization vector with a length of 64 bits. The algorithm uses the following operations: • Addition of 32-bit numbers. • Exclusive OR (XOR). • Bits shift.
The basis of this algorithm is a 64-byte hash
function that works together with a counter and is
20 cycles performed on the internal state. The
disadvantage of Salsa20 is that it can only be used
to protect personal data stored on a PC or hard
disk. Since the integrity of the encrypted text is
not checked, therefore, for more important data, it
is necessary to use authenticated encryption [
The main transformations of the PANAMA cipher operate on 32-bit words. This algorithm can be used for hashing large information arrays. When used as a stream cipher and generator of pseudorandom sequences, the algorithm has a rather long initialization procedure. The field of use of the algorithm is the encryption of video information, for example, in the field of pay TV in this field, where the intensity of the data flow is very high and a high-performance processor is used for its processing, an algorithm that uses the already excessively loaded processor to the smallest extent is needed. The algorithm itself is based on a 544-bit state register and an 8192-bit buffer register. The state of the state register is updated using parallel nonlinear transformations. The buffer register is an LFSR, which is similar to the register used in the SHA hashing algorithm. The state register consists of seventeen 32-bit words. The state of the registers can be changed using two iterations:
• Push iteration accepts input data but does not generate output data.
• Pull iteration does not accept input data but generates output.
There is also a Blank Pull iteration, which is
similar to the Pull iteration, but the output data is
discarded [
HC-256 is a stream encryption algorithm
developed by Wu Hongjun, a cryptographer from
the Singapore Institute of Information and
Communications Research and first published in
2004. A 128-bit version of the cipher was
presented at the aforementioned eSTREAM
competition. The algorithm became one of the
four finalists of the competition in one of the
categories. This algorithm generates a 2128-bit
key sequence using a 256-bit key and a 256-bit
initialization vector. The cipher contains two
secret tables, each of which has 1024 32-bit
elements. At each step, one element from the table
is updated using a nonlinear feedback function,
and after every 2048 steps, all elements of the two
tables will be updated. Such operations as bitwise
exclusive OR, concatenation, shift to the
left/right, and cyclic shift to the right are also used
[
A symmetric block encryption algorithm, a
finalist (winner) of the AES competition and
adopted as an American encryption standard by
the US government [
A symmetric encryption algorithm that was the
US encryption standard from 1976 to the end of
the 1990s and over time gained international use
in various countries around the world. Even from
the time of its development, the algorithm caused
mixed reviews, as it contained classified elements
of its structure - there were fears about the
possibility of control by the US National Security
Agency. The algorithm encrypts data in blocks of
64 bits, and the key length is 56 bits. The
following operations are used in the encryption
process: shuffling, substitution, XOR, key
expansion, and cyclic shift. Currently, DES is
considered unreliable mainly due to the small key
length (56 bits) and block size (64 bits) [
A block symmetric cipher that applies the
mentioned DES algorithm three times to each
block of data. It was created in 1978 based on the
DES algorithm to eliminate the main drawback of
the latter - the short length of the key (56 bits),
which can be broken (chosen) by the sorting
method. The 3DES algorithm works 3 times
slower than DES, but its crypto resistance is much
greater – the time required for 3DES cryptanalysis
is 109 times greater than for DES. The length of
the key is 192 bits, but it is 168 bits long, because
as in DES, in which a 64-bit key is divided into 8
bytes, only 7 bits are used in each byte, so the key
length is 56 bits. Similar to DES, the encryption
process uses operations such as shuffling,
substitution, XOR, key expansion, and cyclic
shift. The main advantage of the algorithm is its
high crypto resistance, however, it has a low speed
of data encryption [
The algorithm was developed in the late 1980s and is the property of RSA Data Security. The development of the algorithm was initiated and partially sponsored by Lotus, which needed a robust encryption algorithm for use in the Lotus Notes system. The stability of the algorithm was checked by the US National Security Agency, certain recommendations were also formulated and implemented by the developers. Encryption takes place in blocks of 64 bits using keys of variable size: from 8 to 1024 bits inclusive (the recommended key size is 64 bits). The RC2 algorithm is a Feistel network, in which 18 rounds of transformations are performed, which are divided into 2 types: • Mixed rounds (mix). • Meshed rounds (mesh).
Also, the following operations are used in the
algorithm: bitwise logical operation AND, bitwise
complement to x, cyclic shift to the left by the
number of bits determined by the substitution
table, and key expansion procedure. According to
studies of the impact of differential and linear
cryptanalysis on the algorithm, the following
result was obtained: the algorithm is not
vulnerable to an attack by the method of linear
cryptanalysis, but it can theoretically be revealed
by the method of differential cryptanalysis [
Block encryption algorithm, which was one of
the finalists of the second stage of the AES
competition. The size of the block is 128 bits, the
length of the key can be 128/192/256 bits, and the
algorithm itself has 32 rounds (16 were initially
planned, but to counter unknown methods of
cryptanalysis, it was increased to 32). Serpent is
an SP network, meaning it has permutation tables
as well as permutation tables. The algorithm has
key expansion operations, linear transformations,
as well as inverse linear transformations (for
decryption). During the development and analysis
of the Serpent algorithm, no vulnerabilities were
found in the full 32-round version (as well as in
the other finalist algorithms). According to the
authors of the algorithm, a new mathematical
theory is needed to break the cipher [
A block symmetric algorithm developed by
B. Schneier in 1993, is not patented and freely
distributed. The length of the key can be from 32
to 448 bits, and the length of the block is 32 bits.
The algorithm is a Feistel network and,
accordingly, includes such operations as XOR,
substitution tables, and addition [
A symmetric block encryption algorithm
developed by a group of specialists led by B.
Schneier, who was also (like Serpent) among the
five finalists of the second stage of the AES
competition. The algorithm is developed based on
Blowfish, SAFER, and Square, the block size is
128 bits, and the key length is 256 bits with the
number of rounds being 16. One of its features is
permutation tables, which are formed depending
on the key. The algorithm itself was implemented
as a mixed Feistel network with 4 branches that
modify each other using Hadamard
cryptotransformations. The algorithm includes
such functions as a cyclic shift by 1 bit, as well as
a whitening function. The study of Twofish with
a reduced number of rounds showed that the
algorithm has a large margin of stability, and
compared to the other finalists of the AES
competition, it turned out to be the most stable.
However, its unusual structure and relative
complexity raised some doubts about the quality
of this stability—this is exactly what played
against it at the AES competition [
A block symmetric algorithm developed by IBM Corporation. According to the results of the AES competition, MARS also reached the finals but lost to Rijndael. MARS is currently distributed under a royalty-free license. The block size in the algorithm is 128 bits, and the key size can be from 128 to 448 bits (must be a multiple of 32 bits). In the encryption process, the algorithm uses the following operations: addition/subtraction, exclusive OR, substitution tables, fixed cyclic shift, and data-dependent cyclic shift, multiplication modulo 232, key expansion. According to IBM, the company’s 25-year cryptanalytic experience is invested in the MARS algorithm and, along with high cryptographic stability, the cipher allows effective implementation even within such limited frameworks as is characteristic of smart cards, which allows it to be used in UAVs. From the point of view of cryptanalysis, there are currently no effective attacks on this algorithm, but it has several weaknesses, in particular:
• Subkeys with a large number of repeated zeros or ones can lead to effective attacks on MARS since weak subkeys will be generated based on them.
• The two least significant bits used in
multiplication are always equal to one, that is, two
input bits are unchanged during the process of
multiplication by the key, as well as two output
bits that are independent of the key [
Developed by RSA Data Security, RC6 is an
evolutionary improvement over its predecessor,
RC5. It was designed specifically for the AES
competition and was one of the finalists. The
algorithm operates on data blocks of 128 bits and
supports key sizes of 128, 192, and 256 bits. RC6
is a parameterized algorithm where the block size,
key size, and number of rounds are adjustable.
The primary innovation of RC6 is the introduction
of integer multiplication as an additional
operation. The algorithm includes operations like
data-dependent rotations, modular addition, and
bitwise XOR. The use of multiplication aims to
provide additional diffusion over RC5 and to
confound linear and differential cryptanalysis
[
GOST 28147-89 is a Soviet and Russian
government standard symmetric key block cipher.
Developed in the 1970s, the standard had been
marked “Top Secret” and then downgraded to
“Secret” in 1990. It was a Soviet alternative to the
United States’ DES. The algorithm operates on
64-bit data blocks with a key length of 256 bits.
It's a Feistel network of 32 rounds. The key
schedule is simple, and the S-boxes (substitution
boxes) were classified, with a few different sets
known to exist. In the context of the standard, the
algorithm was usually called Magma. The cipher
has been used in various Russian state standards
(STBs) and is used in some Russian cryptographic
systems [
Kalyna is a symmetric block cipher that was
selected as the new encryption standard of
Ukraine. The algorithm was designed to have a
high level of resistance against known
cryptanalytic attacks and to achieve a high-speed
performance on contemporary processors. Kalyna
supports block sizes of 128, 256, or 512 bits and
key sizes that can be either equal to or double the
block size. The cipher's structure is based on a
substitution-permutation network (SPN) and
includes operations like SubBytes, ShiftRows,
MixColumns, and AddRoundKey, similar to
AES. However, Kalyna uses different S-boxes
and has a more complex key schedule. The
development of Kalyna was part of a larger effort
to develop cryptographic standards in Ukraine
that would be free from foreign patents and would
meet the modern international cryptographic
strength criteria [
The algorithm itself uses transformation
operations, substitution, and permutation tables,
key expansion, addition modulo 2, linear
transformations, and pre- and post-whitening
[
Artificial neural networks (ANNs) are
considered tools that can help analyze causal
relationships in complex systems [
Consider the potential of training a deep neural
network (DNN) with multiple hidden layers, akin
to the human neural system. The efficacy of the
prior generation of neural networks is primarily
confined to SNNs with one or two hidden layers.
This is because training DNNs poses challenges,
often resulting in a final accuracy that is inferior
to SNNs [
The advantages of employing DNNs with
limited datasets for cryptographic algorithm
determination are evident. Large regression or
classification challenges, previously addressed
using conventional machine learning techniques
(like SNN, SVM, etc.) with small datasets, can
now be tackled using DNNs, yielding superior
accuracy and enhanced generalization [
• Data set pre-processing and partitioning into training/testing subsets.
• Training of SVM, SNN, and DNN.
• Evaluation of the trained machine learning models.
• Model comparison.
• Determining accuracy and selecting the best model.
• Employing the chosen model for predictions.
The DNN training process is bifurcated into
pre-training and fine-tuning phases. Our analysis
will also involve training and evaluating a support
vector machine and a shallow neural network to
validate DNN’s superior accuracy. Post the
training of SVM, SNN, and DNN, a summary of
their training/testing accuracies will be drawn and
juxtaposed. The PCA pre-processing adversely
impacts both training and testing accuracies,
potentially due to the loss of non-linear data. The
subsequent segments of this research will rely on
an unprocessed dataset. A fully connected DNN,
encompassing three or more hidden layers,
demonstrates its superiority over shallow ANNs
and support vector machines by achieving
enhanced prediction accuracy and generalization
[
To facilitate experimental studies, a dedicated library was crafted, encompassing all the aforementioned algorithms in a software format.
The development leveraged the Python programming language, and the Pycryptodome library was employed to instantiate encryption algorithms. This library boasts pre-compiled encryption algorithms available in the .pyc format. The software manifests as a commandline utility, enabling users to designate files for encryption, select the desired algorithm, and determine its operational mode.
The utility operates as follows: 1. Before initiating any task, users must activate the virtual environment to access algorithms from the Pycryptodome library.
2. If no arguments are provided during utility execution, or if the utility is invoked with the -h flag, users are presented with a concise guide on utility usage in the format below. python main.py -[enc/dec] [Enc/Dec method] [Mode] [Key] [InFile] [OutFile] To see supported methods use -> python main.py -s To see supported modes use -> python main.py -m where [enc/dec] denotes encryption or decryption, [Enc/Dec method] specifies the chosen algorithm, [Mode] indicates the operational mode, [Key] represents the encryption key, [InFile] is the source file, and [OutFile] is the destination file.
Additionally, users can invoke the utility with the -s or -m flags to peruse available algorithms and operational modes, respectively.
3. Upon successful encryption or decryption, the resultant files are located in the program's root directory.
To streamline the software tools' implementation and facilitate future enhancements, abstract classes were devised and housed in the Libs/CryptoAbstract.py directory. These classes ensure uniformity across the cryptographic algorithms, guaranteeing a consistent data input and output interface. Furthermore, the Libs/env.py path contains a file that enumerates dictionaries detailing available algorithms and their respective operational modes.
Each algorithm and its various versions are stored in a directory with the appropriate name, and implemented as a class with the following methods: • Initialization. • Block encryption/decryption.
• Reading the file byte by byte for further data encryption.
The software implementation of the dataset is created in such a way that it can be modified and supplemented rather quickly in the future.
For the experiment, files of various sizes (from several kilobytes to several gigabytes) were encrypted. Each file was encrypted by each algorithm up to 10 times (to increase the accuracy of the study).
Experiments were conducted on the following hardware platform: processor—Intel core i7 (9th gen); video card – Nvidia Geforce GTX 1060; RAM—16GB RAM DDR3; operating system – Windows 10.
The results of the experiment were processed and presented in the dataset in the form of a quantitative assessment—a comparison table (Table 2) was created with expert assessments of parameters: • Data on cryptographic resistance (CRS). • Cryptographic resistance reserve (FCRS). • Encryption speed (ECR). • Extension key (KEA).
The first two parameters were taken from open sources, others were verified by experimental research by the authors.
In Table 2 numbers from 1 to 10 determine the effectiveness of the algorithm according to the specified criterion (1 is the worst score, and 10 is the best).
Therefore, a dataset of crypto-algorithms was created, which can be used to ensure the confidentiality of data during transmission from UAVs. This dataset can also be used by ANN to select one or another encryption algorithm depending on the given requirements. Next, the authors plan to select the criteria for the application of this ANN dataset, after which a knowledge base for ANN training will be created.
In the future, a method of increasing the productivity of ANNs will also be developed.
The research has successfully culminated in the creation of an open dataset of cryptographic algorithms. This dataset is designed to bolster the efficacy of information protection during UAV transmissions, leveraging the capabilities of ANNs. As of now, the dataset encompasses a variety of block and stream cryptographic algorithms, detailed within this article. The robustness and performance speed of these algorithms has been ascertained based on extensive scientific investigations conducted by diverse researchers.
Future endeavors aim to refine the dataset by incorporating a broader spectrum of algorithms, including proprietary ones developed by the authors. Efforts will also be directed towards optimizing the software rendition of these algorithms and undertaking supplementary experiments. These experiments will be geared towards curating a comprehensive repository of algorithms tailored for UAV applications, especially in synergy with ANNs. Furthermore, the research ambit will be expanded to explore other challenges associated with UAV operations that can be addressed using the advanced tools of ANN.
The work was carried out as part of the research project “Intellectualized System of Secure Transmission of Packet Data based on Reconnaissance and Search Unmanned Aerial Vehicle” (#0122U002361), financed by the Ministry of Education and Science of Ukraine from 2022–2024.