فصلنامه پدافند الکترونیکی و سایبری
سال نهم شماره 4 (پیاپی 36، زمستان 1400)

Semantic segmentation is one of the most common outputs for vision-based autonomous vehicles. Deep neural networks need massive amount of data in order to learn new environment features with different domain. However, mentioned data always take too much time for humans to manually annotate. In this paper, we use teacher-student technique for deep models interaction. First, we train DABNet and ContextNet as teachers with BDD100K database. Also, we seek to analyze the generalization and robustness of teacher networks by evaluation on CARLA simulator. Finally, teacher networks train FastSCNN model automatically using Cityscapes database without any human interference. Results show that student’s performance in classes like vehicles, people and road which are probably the highest priority classes to detect, has only 1.2%, 3% and 3.8% accuracy difference respectively. Also, there is 4.5% drop for model’s mean intersection over union accuracy between teacher’s performance in comparison to the similar model which is trained on manual annotations.

Today, due to highly advanced attacks and intrusions, it has become very difficult to detect IoT attacks in cloud environments. Other problems with cloud systems include low error detection accuracy, false positive rates, and long computation times. In the proposed method, we present a hybrid intrusion detection model including a clustering algorithm and a machine-based random forest classification for fog and cloud environments. Also, to control the network traffic in the physical layer and also to detect anomalies between IoT devices, calculations will be performed on the fog and the edges of the cloud, so that after preprocessing, the incoming traffic to the fog and cloud will be checked and if necessary They are directed to an anomaly detection module. A random forest-based learning classification was used to identify the type of each attack. General data and cloud data have been used for research. The overall accuracy of the proposed intrusion detection system was 98.03 and the mean false positive rate was 17% and the anomaly detection rate was 96.30, which is considerable compared to previous methods.

Gait is a biometric feature that can be used to identify individuals from videos. Two main challenges are a change in the direction and angle of walking and a change in the appearance due to various reasons such as carrying a bag or any change in clothes that significantly affect identification. In the present paper, a method is proposed which addresses both challenges. In the proposed method, first, the direction of walking is determined using the position of several pixels in the foot zone of the gait energy image (GEI). The pixels are selected to have the maximum identification percentage. Then, the genetic algorithm is used to identify and mask zones of GEI with the most changes in both carrying a bag and changing clothes. The GA is capable of identifying and removing the optimized zone with a good precision that makes the system robust when there are changes in the appearance. Logically, the system should have a good performance because the walking direction identification stage is designed in an affordable computing time. Moreover, GA maintains more useful data comparing to similar techniques. According to the results, an average identification percentage of 95.9% was achieved which confirms the superiority of the proposed method over similar ones.

Natural language generation systems are the subset of natural language processing, have been around for a long time, but their technology as a commercial tool has recently become widespread. In natural language generation, the system needs to decide about how to put a concept among words. The ability for generating a meaningful text plays a key role in many natural language processing applications. The aim of this paper is to propose a method for generating text using artificial intelligence methods with the correct structure, a starting point for generating Persian(Farsi) texts. In order to promote the text generation, it has been attempted to use the combination of machine learning methods and probabilistic models. In the proposed model, the probabilistic models and Word2vec, as a word embedding method, are used to extract the rules and to vectorize the text, respectively. Then, combinating these and the cosine distance are used in the generation phase. The results indicate the performance of proposed model and the generated text has the appropriate structure, concept and variety. Also, the model is optimal in terms of humanity and complexity rather than other methods.

One of the destructive factors in telecommunication and radar systems is intentional interference. Intentional interference is created by using jammer to destroy the enemy's communication and radar systems. If the intentional interference is not reduced well, the efficiency of the telecommunication system will be completely disrupted. Jammers purposefully interfere and affect the optimal performance of the system. The NLMS adaptive algorithm is one of the effective algorithms in eliminating intentional interference. In this paper, a new algorithm for eliminating intentional interference in cognitive radio systems using wavelet transform is presented. In the simulations, a 25-user cognitive radio system (as a victim network) is used in the vicinity of a network of primary users with Markov channel functionality. Considering eleven different scenarios, the performance of the proposed algorithm is investigated. To evaluate the performance of the proposed algorithm, the criterion of successful transmission of information in terms of signal to jammer ratio in each of the scenarios is discussed. According to the simulation results, the proposed algorithm, compared to the adaptive algorithm (NLMS), shows a significant improvement. According to the results, the proposed algorithm shows a 13% improvement in successful transmission at SJR=5dB compared to the NLMS adaptive algorithm

In this paper, we investigate joint information theoretic security and covert communication on a network in the presence of a single transmitter (Alice), a friendly jammer, a single untrusted user, two legitimate users, and a single warden of channel (Willie). In the considered network, one of the authorized users, Bob, needs a secure and covert communication, and therefore his message must be sent securely, and at the same time, the existence of his communication with the transmitter should not be detected by channel’s warden, Willie, Meanwhile, another authorized user, Carol, needs covert communication. The purpose of secure communication is to prevent the message being decoded by the untrusted user who is present on the network, which leads us to use one of the physical layer security methods, named the secure transmission of information theory. In some cases, in addition to protecting the content of the message, it is important for the user that, the existence of the transmission not being detected by an adversary, which leads us to covert communication. But as we know, the main problem in covert communication is the low transmission rate, because we have to reduce the transmission power so that the main message get hide in background noise. In this study, in addition to using joint security of information theory and covert communication, we want to examine the average transmission rate according to the limitations and requirements of covert communication and quality of service requested by users by using the one of power methods, i.e. creating artificial noise in the network by a friendly jammer.

Passive millimeter wave imaging is using to discover hidden objects under human clothes. Discovering hidden objects in the places such as airports, due to their security, is extremely important. However, millimeter wave images have low-quality and image processing techniques are needed to improve the quality of the images. This paper attempts to present a method of fusion approach to discover hidden objects from PMMW images and preserve the detail of visible images. In the proposed method, images are subdivided using BEMD conversion into high frequency and low frequency sub-images. In the next step, the NSST conversion is used to parse images from the previous step in different resolutions and directions, and then the improved SCM neural network is used as the fusion rule. The results are evaluated using fusion effectiveness criteria of QAB/F and MI. Simulation results show that the proposed method improves the best previous results, which were combined using NSST analysis method and ISCM law, with an average of about 33% for the QAB/F criterion.

The security is the key challenge to be well addressed in case of emerging technologies such as the Internet of Things, Internet of Transportation, e-Health, etc. Ignoring this challenge can sometimes cause irreparable personal and financial damage to human beings in everyday life. On the other hand, to identify and extract security requirements and potential threats in the design phase of large-scale and interactive systems, there is a need to model the threats. The problem is that the existing modelling methods are mostly manual, which is associated with error, cost, time consumption, and failure to evaluate all conceivable possibilities. The present paper proposes a new method, called Automated Security Evaluation of Threat Paths, as an automated solution to the problem of identifying and extracting potential threats. In the proposed method, by adding new capabilities such as conditional probability and security to Petri Nets, it is possible not only to automatically generate the threat paths, but also to automatically evaluate the security of threat models in both quantitative and qualitative ways. In this paper, the performance of the proposed method was evaluated under different security scenarios, and the results showed that, compared to other existing methods, the proposed method offers a higher level of security assurance; additionally, it is, unlike the existing methods, fully automated.

The GMCR model is one of the approaches to modeling and analyzing real-world conflicts based on game theory. in this model, as the number of players’ options increases, the number of game states (problem state space) increases exponentially. As the number of game feasible states increases, so does the number of game equilibrium states. Extracting favorable equilibrium states and effective options is one of the requirements of widespread conflicts such as malware games and countermeasures. In this paper, based on the GMCR, the MAG architecture with four processing layers is presented. Mag's architecture was evaluated and analyzed based on methods of detecting and analyzing Detection-Independent and deductive evidence of malware and countermeasures in the form of three related games. The evaluation results showed that among the attacker options, the option "Fileless Cyber Attacks" and among the defense options, the options of "network communication disconnection", "path exploration techniques" and "symbolic execution", at a rate of 100%, are the effective options of the actors. Reducing the game state space by using the game abstraction algorithm, scenario-based and repeated games, extracting effective actions and favorable equilibrium states of the players are some of the advantages of MAG architecture. The MAG architecture can be used in cyber operations decision support systems and tabletop cyber wargames to make the right decisions and respond appropriately.

In recent years, the violation of the privacy of information and communication networks, which are more commonly known as cyber attacks, has been increasing exponentially. These network breaches range from direct attacks on government infrastructure to populist activism and theft. This trend has increased the social and political awareness of users. Active cyber defense is a mechanism to protect computer devices, networks and digital devices against cyber attacks and destructive intrusions. The function of active cyber defense is to actively seek to infiltrate or engage with the perpetrators of cyber incidents. The reaction of the defenders has always caused an imbalance between their activities and the attackers. Attackers have always used the network as a platform to strengthen their attacks to increase the intensity of attacks. In this study, we have used Markov model to model and show the inherent difference in the performance of users and network intruders, with the difference that the proposed model was different with the automatic patch and Moran trend and was more similar to the voter model. The proposed dynamic system model is based on the field average approximation, which purposefully determines the performance of active cyber defense. The imbalance between defenders and attackers in this proposed model is eliminated and according to the proposed dynamics and network conditions, we have provided a platform for the interaction of defenders and attackers to examine the performance of power functions in different situations. In the simulation, the performance of attackers and defenders was examined and according to the results obtained in the diagrams mentioned in the final part of the article, it was shown how to use active cyber defense at the right time so that we can use this defensive approach in the right situation.

Weaknesses in video surveillance security systems have led to attempts to use hybrid methods to address these issues. In this article, we consider the subject of face recognition based on the evaluation of 3D structure and shape. Evaluating and fitting the structure and achieving a 3D shape makes face recognition possible with a wide range of parameters and under different conditions, and intrinsic and extrinsic models are well isolated from the parameters. Exposure conditions and head position are all taken into account and well-controlled. As we know, various parameters such as the angle of the head in front of the camera, the direction of imaging, the amount of exposure, and the types of noise in the image are among the most difficulties and issues and affecting the successful detection rate of human face recognition systems. In the field of face recognition, most of the present methods due to face graph extraction and two-dimensional flat structures by changing the angle of the head, especially in the case of profile or three faces, suffer from errors and incapability. The 3D structure of the face overcomes many of these issues. This method is a combination of geometry tools, statistical meters, and dimensionality reduction methods simultaneously. Accuracy and efficiency testing are based on the correct recognition rate and head position (rotation).

A modern architecture called fog computing is used in IoT-based network systems. Providing data services is economical and low latent in fog computing architecture. This paper addresses the main challenge of allocating computing resources in fog computing. Solving the resource allocation challenge leads to the increased profits, economic savings, and optimal use of the computing systems. In this survey, resource allocation has been improved by using the combined Nash equilibrium algorithm and the auction algorithm. In the proposed method, each player is assigned a specific matrix. Each player’s matrix includes fog nodes, data service subscribers, and data service operators. At each stage of the algorithm, each player generates the best strategy based on the strategy of the other players. The results show the superiority of fog node utility and data service operator utility in the proposed method compared with the Stackelberg game algorithm. The first comparison is based on the changes of subscribers in which the productivity of the node with 240 used subscribers in the proposed method is 6852.8 and it is 5510.2 in the Stackelberg method with the same conditions. The second comparison is based on the service rate of the resource control blocks (μ) in which the productivity of the data service operator with μ=4 in the proposed method is 1.35E + 07 and it is 1E + 7 in the Stackelberg method with the same conditions.