Anomalous Inputs Research Articles

FetalAI: A deep learning web-based application for predicting birthweight from prenatal ultrasound measurements

FetalAI is a web-based application that utilizes a longitudinal learning procedure that integrates data augmentation and deep learning as back-end, with a purpose of predicting an newborn’s projected birthweights and the associated risks of excessive fetal growth conditions, including large-for-gestational-age (LGA) and macrosomia. This article provides an end-to-end overview of FetalAI’s development process, which includes data sourcing and preprocessing, back-end model design and training, application output interpretation, and user interactions. We begin by describing the original training data sources and their mathematical structures. We then outline the data engineering pipeline that acquires and pre-processes user input. After data sourcing and pre-processing, we provide details on the back-end model, which applies mixed-effects models for data augmentation and pre-trained recurrent neural networks (RNNs) to forecast birthweight sequences. After back-end, we discuss the front-end development of FetalAI, with an emphasis on creating the graphical user interface (GUI) using Streamlit and the application’s ability to handle corner cases and anomalous inputs. Our article concludes by highlighting the key functionalities and potential use cases of FetalAI, as well as outlining future directions for research and development in this domain. FetalAI represents a significant step forward in the application of deep learning techniques to improve prenatal care and support clinical decision-making.

Targeted Collapse Regularized Autoencoder for Anomaly Detection: Black Hole at the Center.

Autoencoders have been extensively used in the development of recent anomaly detection techniques. The premise of their application is based on the notion that after training the autoencoder on normal training data, anomalous inputs will exhibit a significant reconstruction error. Consequently, this enables a clear differentiation between normal and anomalous samples. In practice, however, it is observed that autoencoders can generalize beyond the normal class and achieve a small reconstruction error on some of the anomalous samples. To improve the performance, various techniques propose additional components and more sophisticated training procedures. In this work, we propose a remarkably straightforward alternative: instead of adding neural network components, involved computations, and cumbersome training, we complement the reconstruction loss with a computationally light term that regulates the norm of representations in the latent space. The simplicity of our approach minimizes the requirement for hyperparameter tuning and customization for new applications which, paired with its permissive data modality constraint, enhances the potential for successful adoption across a broad range of applications. We test the method on various visual and tabular benchmarks and demonstrate that the technique matches and frequently outperforms more complex alternatives. We further demonstrate that implementing this idea in the context of state-of-the-art methods can further improve their performance. We also provide a theoretical analysis and numerical simulations that help demonstrate the underlying process that unfolds during training and how it helps with anomaly detection. This mitigates the black-box nature of autoencoder-based anomaly detection algorithms and offers an avenue for further investigation of advantages, fail cases, and potential new directions.

Beyond vanilla: Improved autoencoder-based ensemble in-vehicle intrusion detection system

Modern automobiles are equipped with a large number of electronic control units (ECUs) to provide safe, driver assistance and comfortable services. The controller area network (CAN) provides near real-time data transmission between ECUs with adequate reliability for in-vehicle communication. However, the lack of security measures such as authentication and encryption makes the CAN bus vulnerable to cyberattacks, which affect the safety of passengers and the surrounding environment. Detecting attacks on the CAN bus, particularly masquerade attacks, presents significant challenges. It necessitates an intrusion detection system (IDS) that effectively utilizes both CAN ID and payload data to ensure thorough detection and protection against a wide range of attacks, all while operating within the constraints of limited computing resources. This paper introduces an ensemble IDS that combines a gated recurrent unit (GRU) network and a novel autoencoder (AE) model to identify cyberattacks on the CAN bus. AEs are expected to produce higher reconstruction errors for anomalous inputs, making them suitable for anomaly detection. However, vanilla AE models often suffer from overgeneralization, reconstructing anomalies without significant errors, resulting in many false negatives. To address this issue, this paper proposes a novel AE called Latent AE, which incorporates a shallow AE into the latent space. The Latent AE model utilizes Cramér’s statistic-based feature selection technique and a transformed CAN payload data structure to enhance its efficiency. The proposed ensemble IDS enhances attack detection capabilities by leveraging the best capabilities of independent GRU and Latent AE models, while mitigating the weaknesses associated with each individual model. The evaluation of the IDS on two public datasets, encompassing 13 different attacks, including sophisticated masquerade attacks, demonstrates its superiority over baseline models with near real-time detection latency of 25ms.

Probing the Purview of Neural Networks via Gradient Analysis

We analyze the data-dependent capacity of neural networks and assess the anomaly in inputs from the perspective of networks during inference. The notion of data-dependent capacity allows for analyzing the knowledge base of a model populated by learned features from training data. We define <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">purview</i> as the additional capacity necessary to characterize inference samples that differ from training data. To probe the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">purview</i> of a network, we utilize gradients to measure the amount of change required for the model to characterize given inputs more accurately. To eliminate the dependency on ground truth labels in generating gradients, we introduce <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">confounding labels</i> that are formulated by combining multiple categorical labels.We show that our gradient-based approach can effectively differentiate inputs that cannot be represented accurately with learned features. We utilize our approach in applications of detecting anomalous inputs, including out-of-distribution, adversarial, and corrupted samples. Our approach requires no hyperparameter tuning or additional data processing and outperforms state-of-the-art methods by up to 2.7%, 19.8%, and 35.6% of AUROC scores, respectively.

A Global Model Approach to Robust Few-Shot SAR Automatic Target Recognition

In real-world scenarios, it may not always be possible to collect hundreds of labeled samples per class for training deep learning-based SAR Automatic Target Recognition (ATR) models. This work specifically tackles the few-shot SAR ATR problem, where only a handful of labeled samples may be available to support the task of interest. Our approach is composed of two stages. In the first, a global representation model is trained via self-supervised learning on a large pool of diverse and unlabeled SAR data. In the second stage, the global model is used as a fixed feature extractor and a classifier is trained to partition the feature space given the few-shot support samples, while simultaneously being calibrated to detect anomalous inputs. Unlike competing approaches which require a pristine labeled dataset for pretraining via meta-learning, our approach learns highly transferable features from unlabeled data that have little-to-no relation to the downstream task. We evaluate our method in standard and extended MSTAR operating conditions and find it to achieve high accuracy and robust out-of-distribution detection in many different few-shot settings. Our results are particularly significant because they show the merit of a global model approach to SAR ATR, which makes minimal assumptions, and provides many axes for extendability.

Intercity Rail Transit Platform Anomaly Detection Using Door Tracking-Based Key Frame Extraction and AnoDet Network

The anomalies in the risk space between platform doors and train will threaten the safety of intercity rail transit. However, performing anomaly detection tasks with an accurate grasp of the time gap between the train doors closed and the train departure is challenging. The scarcity of anomalous samples and the lack of prior knowledge of anomalies also pose considerable obstacles to anomaly detection. To solve the aforementioned problems, a method based on key frames extraction and a two-stage Transformer network is proposed. The switch statuses of the train door are extracted by the door tracking algorithm, which effectively provides key frames for subsequent anomaly detection. An end-to-end two-stage Transformer, which is named AnoDet, is proposed for unsupervised anomaly localization and supervised multi-classification. The AnoDet is based on an image masking strategy, dual-channel network, knowledge migration base and anomaly object classification set prediction loss. The unsupervised localization model first utilizes the masking strategy to remove a large number of image patches that may contain anomalies in the inputs. Then, it trains a dual-channel network based on the remaining normal patches to inpaint anomalous inputs into normal images. The supervised classifier proposes a knowledge migration base to provide prior knowledge of anomalies and alleviate the scarcity of anomalous samples. It also introduces a set prediction loss to achieve a natural connection with the unsupervised model. In the intercity rail station anomaly dataset, our method achieves 97.8% precision and 96.6% recall for anomaly localization and obtains 88.9% mAP for multi-classification, outperforming the state-of-the-arts.

Quality Criteria and Method of Synthesis for Adversarial Attack-Resistant Classifiers

The actual problem of adversarial attacks on classifiers, mainly implemented using deep neural networks, is considered. This problem is analyzed with a generalization to the case of any classifiers synthesized by machine learning methods. The imperfection of generally accepted criteria for assessing the quality of classifiers, including those used to confirm the effectiveness of protection measures against adversarial attacks, is noted. The reason for the appearance of adversarial examples and other errors of classifiers based on machine learning is investigated. A method for modeling adversarial attacks with a demonstration of the main effects observed during the attack is proposed. It is noted that it is necessary to develop quality criteria for classifiers in terms of potential susceptibility to adversarial attacks. To assess resistance to adversarial attacks, it is proposed to use the multidimensional EDCAP criterion (Excess, Deficit, Coating, Approx, Pref). We also propose a method for synthesizing a new EnAE (Ensemble of Auto-Encoders) multiclass classifier based on an ensemble of quality-controlled one-class classifiers according to EDCAP criteria. The EnAE classification algorithm implements a hard voting approach and can detect anomalous inputs. The proposed criterion, synthesis method and classifier are tested on several data sets with a medium dimension of the feature space.

Anomaly Detection in Multi-Host Environment Based on Federated Hypersphere Classifier

Detecting anomalous inputs is essential in many mission-critical systems in various domains, particularly cybersecurity. In particular, deep neural network-based anomaly detection methods have been successful for anomaly detection tasks with the recent advancements in deep learning technology. Nevertheless, the existing methods have considered somewhat idealized problems where it is enough to learn a single detector based on a single dataset. In this paper, we consider a more practical problem where multiple hosts in an organization collect their input data, while data sharing among the hosts is prohibitive due to security reasons, and only a few of them have experienced abnormal inputs. Furthermore, the data distribution of the hosts can be skewed; for example, a particular type of input can be observed by a limited subset of hosts. We propose the federated hypersphere classifier (FHC), which is a new anomaly detection method based on an improved hypersphere classifier suited for running in the federated learning framework to perform anomaly detection in such an environment. Our experiments with image and network intrusion detection datasets show that our method outperforms the state-of-the-art anomaly detection methods trained in a host-wise fashion by learning a consensus model as if we have accessed the input data from all hosts but without communicating such data.

Improving Out-of-Distribution Detection by Learning From the Deployment Environment

Recognition systems in the remote sensing domain often operate in “open-world” environments, where they must be capable of accurately classifying data from the in-distribution categories while simultaneously detecting and rejecting anomalous/out-of-distribution (OOD) inputs. However, most modern designs use deep neural networks (DNNs) to perform this recognition function that are trained under “closed-world” assumptions in offline-only environments. As a result, by construction, these systems are ill-posed to handle anomalous inputs and have no mechanism for improving OOD detection abilities during deployment. In this work, we address these weaknesses from two aspects. First, we introduce advanced DNN training methods to codesign for accuracy and OOD detection in the offline training phase. We then propose a novel “learn-online” workflow for updating the DNNs during deployment using a small library of carefully collected samples from the operating environment. To show the efficacy of our methods, we consider experimenting with two popular recognition tasks in remote sensing: scene classification in electro-optical satellite images and automatic target recognition in synthetic aperture radar imagery. In both, we find that our two primary design contributions can individually improve detection performance, while also being complementary. Additionally, we find that detection performance on difficult and highly granular OOD samples can be drastically improved using only tens or hundreds of samples collected from the environment. Finally, through analysis, we determine that the logic for adding/removing samples from the collection library is of key importance and using a proper learning rate during the model update step is critical.

VASP: An autoencoder-based approach for multivariate anomaly detection and robust time series prediction with application in motorsport

The aim is to provide a framework for robust time series prediction in the presence of anomalies. The framework is developed based on a data set from motorsport but is not limited to this specific area. In motorsport, the usage of sensors during races is generally restricted. Estimating the outputs of these missing sensors therefore provides an advantage over the competition. Deep learning approaches such as long short-term memory (LSTM) neural networks have proven to be useful for that task, however, their accuracy decreases significantly if anomalies occur in the input signals. To overcome this problem, we propose the variational autoencoder based selective prediction (VASP) framework which combines the tasks of anomaly detection and time series prediction. VASP consists of a variational autoencoder (VAE), an anomaly detector and LSTM predictors. Depending on the anomaly detector, a subset of the inputs may be replaced by the VAE, allowing a more robust prediction. To the best of our knowledge the approach of using a VAE to only selectively replace anomalous input data before prediction has not yet been published. Our contributions are clear implementation guidelines and a comparison to other VAE-based methods and a LSTM approach as baseline. We simulate anomalies with three approaches and show that VASP outperforms other methods by having no trade-off between accuracy and robustness. VASP is as accurate as the baseline for regular data, but for anomalous inputs the error is reduced by 13% to 33% on average and up to 70% in special cases.

Regularizing Attention Networks for Anomaly Detection in Visual Question Answering

For stability and reliability of real-world applications, the robustness of DNNs in unimodal tasks has been evaluated. However, few studies consider abnormal situations that a visual question answering (VQA) model might encounter at test time after deployment in the real-world. In this study, we evaluate the robustness of state-of-the-art VQA models to five different anomalies, including worst-case scenarios, the most frequent scenarios, and the current limitation of VQA models. Different from the results in unimodal tasks, the maximum confidence of answers in VQA models cannot detect anomalous inputs, and post-training of the outputs, such as outlier exposure, is ineffective for VQA models. Thus, we propose an attention-based method, which uses confidence of reasoning between input images and questions and shows much more promising results than the previous methods in unimodal tasks. In addition, we show that a maximum entropy regularization of attention networks can significantly improve the attention-based anomaly detection of the VQA models. Thanks to the simplicity, attention-based anomaly detection and the regularization are model-agnostic methods, which can be used for various cross-modal attentions in the state-of-the-art VQA models. The results imply that cross-modal attention in VQA is important to improve not only VQA accuracy, but also the robustness to various anomalies.

Process recovery by rollback and input modification

An effective countermeasure against software vulnerability attacks is a security system that executes a process in a confined or monitored environment. This type of security system isolates compromised software from other software and thus minimises the influence of attacks. However, several problems occur when a security system, without taking any other action, terminates a process that violates a security policy or encounters a fault. In this paper, we propose ProcHealer, a security system that does not terminate the offending process, but instead rolls it back to the pre-violation or pre-fault state where the process can be restarted and remediated. Security policy violations and faults are often caused by anomalous external inputs. In the restarted execution, ProcHealer therefore provides the process with a modified version of the external input.

Functional organisation of visual pathways in a patient with no optic chiasm

Congenital achiasma offers a rare opportunity to study reorganization and inter-hemispheric communication in the face of anomalous inputs to striate cortex. We report neuroimaging studies of a patient with seesaw nystagmus, achiasma, and full visual fields. The subject underwent structural magnetic resonance imaging (MRI), diffusion tensor imaging (DTI) studies, and functional MRI (fMRI) using monocular stimulation with checkerboards, motion, objects and faces, as well as retinotopic quadrantic mapping. Structural MRI confirmed the absence of an optic chiasm, which was corroborated by DTI tractography. Lack of a functioning decussation was confirmed by fMRI that showed activation of only ipsilateral medial occipital cortex by monocular stimulation. The corpus callosum was normal in size and anterior and posterior commissures were identifiable. In terms of the hierarchy of visual areas, V5 was the lowest level region to be activated binocularly, as were regions in the fusiform gyri responding to faces and objects. The retinotopic organization of striate cortex was studied with quadrantic stimulation. This showed that, in support of recent findings, rather than projecting to an ectopic location contiguous with the normal retinotopic map of the ipsilateral temporal hemi-retina, the nasal hemi-retina's representation overlapped that of the temporal hemi-retina. These findings show that congenital achiasma can be an isolated midline crossing defect, that information transfer does not occur in early occipital cortex but at intermediate and higher levels of the visual hierarchy, and that the functional reorganisation of striate cortex in this condition is consistent with normal axon guidance by a chemoaffinity gradient.

Impact of human activities on the central Mediterranean offshore: Evidence from Hg distribution in box-core sediments from the Ionian Sea

Total Hg concentrations have been measured for five box-core sediments collected seawards of the Augusta industrial area (SE Sicily). In more coastal sediments, upcore increasing Hg concentrations, exceeding the Hg background concentration estimated for the Strait of Sicily, indicate Hg contamination over time due to the industrial area development. Strong correlation between total organic C (TOC) and Hg concentrations was found only for core BX2, that displays organic C to total N (C/N) ratios indicative of autochthonous organic matter. For other sediments, high Hg enrichment factors with respect to TOC indicate, in addition to Hg trapping by TOC, other factors as responsible for Hg accumulation. In the presence of some contribution of detrital organic matter, Hg is mainly adsorbed onto the mineral component of the bottom sediments probably because TOC is saturated by Hg excess. Contaminant impact affected also the open sea environment. Main drivers of Hg flux towards the offshore were dredged materials, which repeatedly discharged sediment, resulting in substantial increases in TOC contents and high C/N ratios. Consistent with the geochemistry of recent turbidites, these anomalous sedimentary inputs induced sediment redox environment modifications, constrained by Mn peaks, which affected Hg distribution.

Trajectory Shifts in the Arctic and Subarctic Freshwater Cycle

Manifold changes in the freshwater cycle of high-latitude lands and oceans have been reported in the past few years. A synthesis of these changes in freshwater sources and in ocean freshwater storage illustrates the complementary and synoptic temporal pattern and magnitude of these changes over the past 50 years. Increasing river discharge anomalies and excess net precipitation on the ocean contributed approximately 20,000 cubic kilometers of fresh water to the Arctic and high-latitude North Atlantic oceans from lows in the 1960s to highs in the 1990s. Sea ice attrition provided another approximately 15,000 cubic kilometers, and glacial melt added approximately 2000 cubic kilometers. The sum of anomalous inputs from these freshwater sources matched the amount and rate at which fresh water accumulated in the North Atlantic during much of the period from 1965 through 1995. The changes in freshwater inputs and ocean storage occurred in conjunction with the amplifying North Atlantic Oscillation and rising air temperatures. Fresh water may now be accumulating in the Arctic Ocean and will likely be exported southward if and when the North Atlantic Oscillation enters into a new high phase.

Olfaction in olfactory bulbectomized rats.

Experimental rats had their right olfactory bulb removed on postnatal day 2 (P2) and their left olfactory bulb removed on P90. Control rats had one or both olfactory bulbs removed on P90. Before and after their adult-stage surgery, rats were trained using olfactometry and operant conditioning to detect and discriminate odors. Anterograde transport of horseradish peroxidase applied to the olfactory epithelium revealed numerous axons of olfactory sensory neurons in the right hemisphere of 27 experimental rats. These axons terminated in glomerular-like clusters within the frontal neocortex (n = 5) or anterior olfactory nucleus with some axons extending into the subventricular epithelium (n = 22). Seventeen of the experimental rats were able to detect a variety of odors and to discriminate between odors. Performance accuracy was related to the location and density of these anomalous inputs; experimental rats with inputs confined to frontal neocortex and those lacking any inputs to the forebrain were anosmic, as were adult-operated bilaterally bulbectomized rats. Our results provide strong support for the contention that, in the absence of the olfactory bulbs, olfactory connections to novel forebrain sites can support both odor detection and odor discrimination.