Deep learning models have demonstrated exceptional performance in medical image classification tasks. However, their susceptibility to adversarial attacks raises significant concerns, especially in critical domains like healthcare. In response, various defense mechanisms have been developed to enhance model robustness. Despite the efforts made to enhance model robustness, many existing methods face limitations, such as high computational costs, degraded image quality, and reduced model accuracy, which motivate the need for more effective solutions. In response, this study examines how resilient neural networks (NNs) are against adversarial attacks when used to classify chest X-ray images. This article addresses how steganography techniques, which are often utilized to embed data imperceptibly, could enhance adversarial perturbations, making attacks less visible and posing a challenge to current defenses. These techniques highlight the need for stronger defenses by making adversarial attacks more difficult for automated systems and human observers to detect by hiding perturbations in images. We investigate the embedding of adversarial perturbations using wavelet, Joint Photographic Experts Group (JPEG), and discrete cosine transform (DCT) methods, followed by the application of discrete Fourier transform (DFT), and Gaussian filtering used as defensive strategies against adversarial attacks. These methods leverage distinct properties of image processing and frequency domain analysis to mitigate the impact of adversarial perturbations. Furthermore, this paper comprehensively analyzes each technique, detailing its effectiveness in preserving image quality while reducing susceptibility to adversarial noise. Our findings demonstrate that combining these defense mechanisms can significantly improve the resilience of machine learning models against a wide range of adversarial strategies.

This article analyzes SplitFed Learning against model poisoning vulnerability and develops methods to protect such a system against these attacks. SplitFed learning is a distributed learning paradigm where a neural network model is split between clients and the server, contrasting with traditional Federated Learning. SplitFed learning enables enhanced security and data privacy, and clients do not need to perform heavy computation in model training, as they only need to train a part of the model. This approach ensures that the model can make precise predictions while maintaining the confidentiality of sensitive information. In addition to implementing a SplitFed model, the paper proposes a distance-based method that can poison SplitFed learning-based systems. Subsequently, this paper develops a novel prevention strategy based on robust statistical properties of the sample. To test the proposed methodology, we used the image cell dataset of the malaria parasite as a test case. By addressing the impacts of adversarial attacks, this paper contributes to the advancement of deep learning techniques.

Deep Learning (DL) technology has become ubiquitous in multiple industrial domains, and continuing research and development in this field persists rapidly. One essential application of DL is object detection (OD), which is used for multiple industrial applications. Unfortunately, DL-enabled OD (DLOD) models can fall victim to adversarial attacks. In this paper, we evaluate the effectiveness of using ensemble learning (EL) as a mechanism to defend DLOD models from adversarial attacks in industrial applications. This investigation is structured as a multi-faceted and multi-staged approach that amalgamates both experimental assessments and in-depth analyses. Results indicate that the usage of EL can prevent substantial deterioration in model performance against adversarial samples. This can be attributed to EL’s function of relying on high-performing base learners that have good generalization ability on unseen instances. Additionally, we also note that across both industrial OD datasets, the adversarial samples developed on ResNet-50 impacted the performance of ResNet-50 in Figure negatively. However, due to EL having the ability to validate perturbed samples on a wider variety of strong learners, the model achieved higher performance compared to solely ResNet-50. Achieved results indicate that using EL for protecting against adversarial attacks has merit and further investigations in this direction are warranted.

Currently, Federated Learning is a research approach where multiple parties can train a model together without sharing each other’s data to solve complex problems in machine learning. Ultrasound Nerve Segmentation is a computer vision technique that automatically identifies and segments nerve structures in ultrasound images. This technique is particularly important in medical applications where accurate localization of nerves is crucial, such as during anesthesia, nerve blocks, or surgical procedures. Ultrasound Nerve Segmentation can help doctors find nerves better during medical procedures. This could make patients feel better and have better results. Talking about surgery can make even brave patients scared because it hurts and can cause a lot of pain afterward. To reduce pain, people use drugs called narcotics, but these drugs can have bad side effects. The goal of this project is to improve pain management by using indwelling catheters that block or reduce pain at its source. These catheters reduce the need for painkillers and hasten patient healing. It is crucial to precisely identify nerve structures in ultrasound images to guarantee the exact insertion of a patient’s pain management catheter. We attempt to build an algorithm that can recognize nerve structures in a dataset of neck ultrasound images in this study. To do this, we created a U-net architecture model that will accept an image as input and forecast an image with the source of the pain highlighted as the output. Achieving this objective would improve catheter placement precision and help in the future with reduced pain.

Multiple datasets have been proposed to create Machine Learning (ML)-based Network Intrusion Detection Systems (NIDS). However, many of these datasets suffer from sub-optimal performance and inadequate tail class representation. In this paper, we propose the University of Nevada - Reno Intrusion Detection Dataset (UNR-IDD), which utilizes network port statistics for fine-grained analysis of intrusions. Evaluation results show that UNR-IDD is better than existing NIDS datasets with an Fμ, score of 94% and a minimum F-score of 86%. This is mainly because of sufficient and equal representation of various anomaly types in the UNR-IDD dataset.

The rapid growth in Internet of Things (IoT) technology has become an integral part of today’s industries forming the Industrial IoT (IIoT) initiative, where industries are leveraging IoT to improve communication and connectivity via emerging solutions like data analytics and cloud computing. Unfortunately, the rapid use of IoT has made it an attractive target for cybercriminals. Therefore, protecting these systems is of utmost importance. In this paper, we propose a federated transfer learning (FTL) approach to perform IIoT network intrusion detection. As part of the research, we also propose a combinational neural network as the centerpiece for performing FTL. The proposed technique splits IoT data between the client and server devices to generate corresponding models, and the weights of the client models are combined to update the server model. Results showcase high performance for the FTL setup between iterations on both the IIoT clients and the server. Additionally, the proposed FTL setup achieves better overall performance than contemporary machine learning algorithms at performing network intrusion detection.

Healthcare is one of the foremost applications of machine learning (ML). Traditionally, ML models are trained by central servers, which aggregate data from various distributed devices to forecast the results for newly generated data. This is a major concern as models can access sensitive user information, which raises privacy concerns. A federated learning (FL) approach can help address this issue: A global model sends its copy to all clients who train these copies, and the clients send the updates (weights) back to it. Over time, the global model improves and becomes more accurate. Data privacy is protected during training, as it is conducted locally on the clients’ devices.However, the global model is susceptible to data poisoning. We develop a privacy-preserving FL technique for a skin cancer dataset and show that the model is prone to data poisoning attacks. Ten clients train the model, but one of them intentionally introduces flipped labels as an attack. This reduces the accuracy of the global model. As the percentage of label flipping increases, there is a noticeable decrease in accuracy. We use a stochastic gradient descent optimization algorithm to find the most optimal accuracy for the model. Although FL can protect user privacy for healthcare diagnostics, it is also vulnerable to data poisoning, which must be addressed.

Malwares have historically been the central threat concerning cybersecurity. With the rise in modern technologies, malwares have become more lethal due to the increasing connectivity between various computing environments. To exacerbate this increasing threat, modern day anti-virus signature matching seems to provide limited effectiveness, which makes it possible that new malware variants can evade detection when their behavior does not correlate exactly with known malware signatures on a system. There is also a wide time gap between when a new malware variant is released, and its corresponding antivirus signature is developed and made available to the public. This time gap is when these new malware variants can cause havoc on computing environments. To counteract this threat, the paper proposes a novel malware detection methodology to analyze and detect malware and classify malware by family. This is performed through time series analysis of the sequence of API calls made, using Extreme Learning Machines (ELM) and Online Sequential Extreme Learning Machines (OS-ELM). The paper further analyzes the approach by accurately detecting and predicting malware family classification using a subset of the API call sequences. Simulation results show that the proposed approach can accurately detect and classify malware families with higher accuracy and greater speed than traditional LSTM. Our experiments show ELM and OS-ELM performing with 91% accuracy within just 3 seconds of learning time in contrast to other methods. Thus, our method is superior both in terms of training time and accuracy.

Malware remains one of the gravest threats to cybersecurity, second only to social engineering or a lack of user security awareness. This is especially true for Windows systems in enterprise environments. As malware continues to evolve and frustrate legacy detection and prevention mechanisms, additional approaches are necessary to ensure security resilience. Machine learning offers many opportunities to better combat malware threats through the advantage of big datasets. Our research highlights how machine learning can be leveraged to identify malware threats with rapid results, enabling cybersecurity professionals to learn and adapt to these threats. The approach we present in this paper produces an efficient methodology to discern malware family and function through analysis of just the first 3,000 Windows system API function calls. We compare MLP, CNN, and SVM networks to determine the best performance in terms of accuracy and speed and find that MLP works the best with our dataset.

Deployment of SDN control plane in high-end servers allow many network applications to be automated and easily managed. In this paper, we propose an SDN anomaly detection application, Confident and Explainable Anomaly Detector (CEAD), that automatically detects malicious network flows in SDN-based network architectures. The proposed application employs a set of Machine Learning (ML) classifiers to improve the confidence score of a prediction, thereby creating improved trust upon the prediction, while providing interpretability to the anomaly detector. The method utilizes the Explainable Artificial Intelligence (XAI) framework to provide interpretation to predictions to unearth network features that establish the most influence between predicted anomaly types. Results show that the proposed framework can achieve efficient anomaly detection performance, with near perfect confidence scores. Analysis with XAI highlights that byte and packet transmissions, and their robust statistics, can be significant indicators for prevalence of any attacks. Results also indicate that a subset of influential features can generally be used to decipher between normal and anomalous flow, while certain dataset features can be specifically influential in detecting specific attack types. This can lead to more efficient network resource utilization.

The presence of anomalies or outliers within time-series data can have a detrimental effect on the efficiency of automated decision-making applications. For example, in the context of vehicular traffic flow, various services reliant on traffic data may be negatively impacted by anomalies. This paper presents an automated anomaly detection method based on supervised Long-Short Term Memory (LSTM) neural network and statistical analysis. We train LSTM neural network to predict non-robust statistical properties and combine them with robust properties to determine the anomalies in time-series data. The proposed method relies on segmentation and tunable parameters for anomaly test. We measure the efficacy of our method in terms of Precision, Recall, and F-measure. The metrics approach to 100% for certain instances. We also analyzed the performance on the prevalence of anomalies and on varying specific parameters of the model.

Proliferation of the connected Internet of things (IoT) devices and applications like augmented reality have resulted in a paradigm shift in computation requirement and power management of these devices. Furthermore, processing enormous amounts of data generated by ubiquitous IoT devices and meeting real-time deadline requirements of novel IoT applications exacerbate the challenges in IoT design. To address these challenges, in this paper, we propose a computation offloading architecture to process the huge amount of data generated by IoT devices while simultaneously meeting the real-time deadlines of IoT applications. In our proposed architecture, a resource-constrained IoT device requests a relatively resourceful computing device (e.g., a personal computer) in the same local network for computation offloading. Additionally, in our proposed computation offloading architecture, both client and server devices tune their tunable parameters, such as operating frequency and number of active cores, to meet the application's real-time deadline requirements. We compare our proposed computation offloading architecture with contemporary computation offloading models that use cloud computing. Experimental results verify that our proposed architecture provides a performance improvement of 21.4% on average as compared to cloud-based computation offloading schemes.

This paper proposes a novel approach to integrate plugged-in electric vehicles (PEVs), electric vehicle supply equipments (EVSEs), and smart grid (SG) infrastructure using state of the art software-defined network (SDN) technology, which has a potential to provide unprecedented flexibility to smart grid communication network. We further present set-cardinality based search algorithms for assigning charging station to PEVs to reduce their average charging time. Simulation results show a considerable improvement in the average charging time of PEVs as compared to conventional minimum distance based charging station selection.