Search

Transfer-based attacks, a type of adversarial attack, have become a growing threat in recent years with the proliferation of cloud services. Deep neural networks that exploit human cognitive bias (Loosely Symmetric-Deep Neural Network, LS-DNN) are known as a defensive technique against transfer-based attacks. LS-DNN can prevent malfunctions caused by adversarial examples with a high probability by incorporating human learning characteristics into the neural network’s nodes. However, maintaining accuracy against normal data and reducing Training Time is challenging. This paper proposes a new model called “LS+-DNN” inspired by the Dropout method to solve this problem. Evaluation experiments on two datasets show that the proposed model can achieve both in high dimensions. In addition, we analyze the proposed model focusing on the variance and confidence score of the training parameters of the proposed model. As a result, we point out that the confidence score is an important indicator of robust models against transfer-based attacks.

Although GPS is commonly used for the autonomous flying of unmanned aerial vehicles (UAVs), researchers mainly focus on image-based localization methods due to their tremendous advantages when it comes to GPS-denied environments. In this study, we study the problem of image-based geo-localization between UAV and satellite (known as cross-view geo-localization), which is an essential step towards image-based localization. In cross-view geo-localization, extracting fine-grained features containing contextual information from images is challenging due to the large gap in visual representations between different views. Existing methods in this field often use convolutional neural networks (CNNs) as feature extractors. However, CNNs have some limitations in receptive fields, which leads to the loss of fine-grained information. Some researchers have implemented Transformer-based networks to overcome these circumstances. However, these approaches only focused on understanding the meaning of each pixel based on their attention and only partially utilized tokens that are produced from Transformer blocks. Therefore, different from these works, we proposed a Vision Transformer-based network that takes advantage of local tokens, especially the classification token. Through experiments, our proposed model has significantly outperformed existing state-of-the-art models, which gave promising capabilities for developing this method in the future.

The best-of-n problem [Valentini G, Ferrante E, Dorigo M. The best-of-n problem in robot swarms: formalization, state of the art, and novel perspectives. Frontiers in Robotics and AI. 2017;4(9):1–18.] is a collective decision-making problem in which many robots (agents) select the best option among a set of n alternatives and is focused on the field of distributed autonomous robotic systems and swarm robotics. It is desirable to develop a collective decision-making algorithm that can work even when there are a lot of social–behavioural alternatives (n>2) to realize an intelligent system that can solve more complicated problems. However, previous studies mainly focused on binary collective decision-making scenarios (n = 2). In this paper, we propose a collective decision-making algorithm for the best-of-n problem with a large number of options by using short-term experience memory with a trial and error approach at the group level. After proposing this decision-making process, we show typical behaviour. Next, we show the convergence of this algorithm when a quadratic function is used for the bias corresponding to the individual characteristic. Next, the bias distribution proposed shows that an equilibrium point where all candidates have the same number of supports is unstable, and consensus states are a stable fixed point. Therefore, dynamics is expected to converge towards a consensus. Simulation results and mathematical analysis show that the average time required to find the best option is nearly proportional to the number of options and does not depend on the number of robots.

Currently, malware is increasing in both number and complexity dramatically. Several techniques and methodologies have been proposed to detect and neutralize malicious software. However, traditional methods based on the signatures or behaviors of malware often require considerable computational time and resources for feature engineering. Recent studies have applied machine learning to the problems of identifying and classifying malware families. Combining many state-of-the-art techniques has become popular but choosing the appropriate combination with high efficiency is still a problem. The classification performance has been significantly improved using complex neural network architectures. However, the more complex the network, the more resources it requires. This paper proposes a novel lightweight architecture by combining small Convolutional Neural Networks and advanced Variational Autoencoder, enhanced by channel and spatial attention mechanisms. We achieve overperformance and sufficient time through various experiments compared to other cutting-edge techniques using unbalanced and balanced Malimg datasets.

Malware is becoming an effective support tool not only for professional hackers but also for amateur ones. Due to the support of free malware generators, anyone can easily create various types of malicious code. The increasing amount of novel malware is a daily global problem. Current machine learning-based methods, especially image-based malware classification approaches, are attracting significant attention because of their accuracy and computational cost. Convolutional Neural Networks are widely applied in malware classification; however, CNN needs a deep architecture and GPUs for parallel processing to achieve high performance. By contrast, a simple model merely contained a Multilayer Perceptron called MLP-mixer with fewer hyperparameters that can run in various environments without GPUs and is not too far behind CNN in terms of performance. In this study, we try applying an Autoencoder (AE) to improve the performance of the MLP-mixer. AE is widely used in several applications as dimensionality reduction to filter out the noise and identify crucial elements of the input data. Taking this advantage from AE, we propose a lightweight ensemble architecture by combining a customizer MLP-mixer and Autoencoder to refine features extracted from the MLP-mixer with the encoder-decoder architecture of the autoencoder. We achieve overperformance through various experiments compared to other cutting-edge techniques using Malimg and Malheur datasets which contain 9939 (25 malware families) and 3133 variant samples (24 malware families).

As the development of technology, the demand of operating UAVs (Unmanned Aerial Vehicle) in group grows larger. In the operation of group UAVs, it is important that who the leaders should be and who the followers will be. We use multi agent technique to this problem. This study proposes the mechanism of leader selection based on influence of an agent to its neighbors' movement. The influence value can be calculated by the forces from neighborhood agents. In the proposed method, the leaders are selected dynamically and adaptively. Simple computer simulation shows the effectiveness of the method.

The ability to stop malware as soon as they start spreading will always play an important role in defending computer systems. It must be a huge benefit for organizations as well as society if intelligent defense systems could themselves detect and prevent new types of malware as soon as they reveal only a tiny amount of samples. An approach introduced in this paper takes advantage of One-shot/Few-shot learning algorithms to solve the malware classification problems using a Memory Augmented Neural Network in combination with the Natural Language Processing techniques such as word2vec, n-gram. We embed the malware’s API calls, which are very valuable sources of information for identifying malware’s behaviors, in the different feature spaces, and then feed them to the one-shot/few-shot learning models. Evaluating the model on the two datasets (FFRI 2017 and APIMDS) shows that the models with different parameters could yield high accuracy on malware classification with only a few samples. For example, on the APIMDS dataset, it was able to guess 78.85% correctly after seeing only nine malware samples and 89.59% after fine-tuning with a few other samples. The results confirmed very good accuracies compared to the other traditional methods, and point to a new area of malware research.