Fixed-length Vector Research Articles

A fisher score-based multi-instance learning method assisted by mixture of factor analysis

Multi-instance learning (MIL) is a problem with significant importance due to the existence of ambiguous labels in many practical tasks. The training data in MIL problem is represented by labeled bags containing a various number of unlabeled instances. This inexact label information has brought complexities to the classification of instances and bags. Fisher scores have been utilized to extract fixed-length representation vectors from the MIL bags with the assistance of generative models. However, the generative model utilized by the existing methods might disregard the covariance information. In this paper, we propose a multi-instance learning method (miMFA) utilizing Fisher scores derived from the mixture of factor analysis (MFA) model. The MFA generates nonlinear data with a set of local factor analysis models, while each local model approximates the full covariance Gaussian using latent factors. Thus, the MFA could cover the data distribution and generate Fisher scores effectively. The MFA-based Fisher score is then utilized to form the bag representation. Moreover, a reinforcement vector and the complementary information from auxiliary MIL methods are introduced for performance refinement. The proposed miMFA is evaluated on 11 MIL datasets, and achieves satisfactory results compared with the existing methods.

PredAoDP: Accurate identification of antioxidant proteins by fusing different descriptors based on evolutionary information with support vector machine

Antioxidant proteins play a vital role in diseases prevention caused by free radical intermediates. Accurate identification of antioxidant proteins may provide significant clues to improve their fundamental understanding in the field of bioinformatics and pharmacology. Prediction of antioxidant protein is meaningful and a challenging task. In this study, we develop a novel predictor called PredAoDP based on two different descriptors, which incorporates salient evolutionary profile features with support vector machine (SVM), to predict antioxidant proteins. The evolutionary information in the PSI-BLAST profile is encoded and transformed into a series of fixed-length feature vectors by extracting 20-D, amino acid composition and 400-D bigram features from position-specific scoring matrix (PSSM). As a result, a feature vector of 420-dimensional (420-D) feature vector via serial combination was constructed from two training datasets of antioxidant protein Z1 and Z2. The descriptors are then fed to the SVM for classification along with jack-knife cross-validation test method for evaluating the prediction performance. Our proposed intelligent predictor achieved ACC, 93.18% and MCC, 0.712 for Z1 similarly for Z2 datasets ACC, 97.89% and MCC, 0.949 by jack-knife cross-validation test. To evaluate the generalization abilities of the developed method, we performed an independent test and achieved superior performance compared the other published approaches. The experimental results revealed the effectiveness of the proposed approach and can be utilized as a reliable tool for predicting large-scale antioxidant proteins in particular and other proteins in general. The source code and benchmark datasets are available at https://github.com/saeed344/PreAoDp.

An End-to-End Deep Learning Method for Dynamic Job Shop Scheduling Problem

Job shop scheduling problem (JSSP) is essential in the production, which can significantly improve production efficiency. Dynamic events such as machine breakdown and job rework frequently occur in smart manufacturing, making the dynamic job shop scheduling problem (DJSSP) methods urgently needed. Existing rule-based and meta-heuristic methods cannot cope with dynamic events in DJSSPs of different sizes in real time. This paper proposes an end-to-end transformer-based deep learning method named spatial pyramid pooling-based transformer (SPP-Transformer), which shows strong generalizability and can be applied to different-sized DJSSPs. The feature extraction module extracts the production environment features that are further compressed into fixed-length vectors by the feature compression module. Then, the action selection module selects the simple priority rule in real time. The experimental results show that the makespan of SPP-Transformer is 11.67% smaller than the average makespan of dispatching rules, meta-heuristic methods, and RL methods, proving that SPP-Transformer realizes effective dynamic scheduling without training different models for different DJSSPs. To the best of our knowledge, SPP-Transformer is the first application of an end-to-end transformer in DJSSP, which not only improves the productivity of industrial scheduling but also provides a paradigm for future research on deep learning in DJSSP.

Hybrid CNN-LSTM Architecture for LiDAR Point Clouds Semantic Segmentation

The semantic segmentation of LiDAR in the outdoor environment is still an open problem in the field of automatic driving. Although the emerging development and technological advancements, it remains challenging for three reasons: 1) uneven distribution of LiDAR Points in 3D space; 2) the same object can be represented by different point clouds sequences; 3) no unique geometric information compared with the pixels of 2D RGB images. In order to tackle all the challenges mentioned above, we propose a new LiDAR point clouds semantic segmentation algorithm, Hybrid CNN-LSTM, which is composed of an efficient point clouds feature processing method and a novel neural network structure. Inspired by representing the point clouds as a fixed-length vector in PolarNet, we first convert the 3D point clouds into pseudo image. In order to better represent the features of small objects, we input the pseudo image into long short-term memory (LSTM) network according to the spatial filling curve. We design a new neural network structure, which combines the features of different channels generated by convolutional neural network and long short-term memory network. Experiments show that our method has higher semantic segmentation accuracy in comparison with state-of-the-art algorithms on SemanticKITTI dataset. Theoretically, we provide an analysis to understand why our network can better segment the small objects with sparse point clouds features.

SOIT: Segmenting Objects with Instance-Aware Transformers

This paper presents an end-to-end instance segmentation framework, termed SOIT, that Segments Objects with Instance-aware Transformers. Inspired by DETR, our method views instance segmentation as a direct set prediction problem and effectively removes the need for many hand-crafted components like RoI cropping, one-to-many label assignment, and non-maximum suppression (NMS). In SOIT, multiple queries are learned to directly reason a set of object embeddings of semantic category, bounding-box location, and pixel-wise mask in parallel under the global image context. The class and bounding-box can be easily embedded by a fixed-length vector. The pixel-wise mask, especially, is embedded by a group of parameters to construct a lightweight instance-aware transformer. Afterward, a full-resolution mask is produced by the instance-aware transformer without involving any RoI-based operation. Overall, SOIT introduces a simple single-stage instance segmentation framework that is both RoI- and NMS-free. Experimental results on the MS COCO dataset demonstrate that SOIT outperforms state-of-the-art instance segmentation approaches significantly. Moreover, the joint learning of multiple tasks in a unified query embedding can also substantially improve the detection performance. Code is available at https://github.com/yuxiaodongHRI/SOIT.

Estimating the degree of conflict in speech by employing Bag-of-Audio-Words and Fisher Vectors

The automatic detection of conflict situations from human speech has several straightforward applications such as the surveillance of public spaces, providing feedback about employees in call centers, and other roles in human–computer interactions. In this study we examine the potential of different state-of-the-art feature extraction techniques, all developed to be able to efficiently represent a variable-length speech utterance by a fixed-length feature vector. Besides the ‘ComParE functionals’ attribute set, which became the de facto standard feature set in the area of computational paralinguistics (which focuses on the automatic assessment of non-verbal phenomena being present in human speech), we experiment with two methods introduced quite recently: Bag-of-Audio-Words (BoAW) and Fisher Vectors (FV). Using three standard basic, low-level feature sets, we found that, while BoAW proved to be quite sensitive to its meta-parameter settings, with Fisher Vectors we were able to achieve state-of-the-art conflict intensity estimation performance on a public and widely-used corpus. Furthermore, by applying Principal Component Analysis on the frame-level attributes, we managed to achieve a 30% speed-up in the feature extraction step. Interestingly, in contrast with our previous paralinguistic studies, combining the different predictions with these feature extraction approaches, we were unable to achieve any further significant improvement. The highest correlation coefficient values we got on the test set lay in the range 0.850–0.860, while the authors of several previous studies were able to achieve similar values (i.e. 0.849, 0.856 and 0.853). Considering that in this task the target score to be estimated (i.e. the intensity of the conflict being present in the actual clip) is definitely prone to subjectivity and therefore to label noise, current efforts have probably achieved the highest correlation coefficients attainable, and match human performance.

Efficient analysis of COVID-19 clinical data using machine learning models.

Because of the rapid spread of COVID-19 to almost every part of the globe, huge volumes of data and case studies have been made available, providing researchers with a unique opportunity to find trends and make discoveries like never before by leveraging such big data. This data is of many different varieties and can be of different levels of veracity, e.g., precise, imprecise, uncertain, and missing, making it challenging to extract meaningful information from such data. Yet, efficient analyses of this continuously growing and evolving COVID-19 data is crucial to inform — often in real-time — the relevant measures needed for controlling, mitigating, and ultimately avoiding viral spread. Applying machine learning-based algorithms to this big data is a natural approach to take to this aim since they can quickly scale to such data and extract the relevant information in the presence of variety and different levels of veracity. This is important for COVID-19 and potential future pandemics in general. In this paper, we design a straightforward encoding of clinical data (on categorical attributes) into a fixed-length feature vector representation and then propose a model that first performs efficient feature selection from such representation. We apply this approach to two clinical datasets of the COVID-19 patients and then apply different machine learning algorithms downstream for classification purposes. We show that with the efficient feature selection algorithm, we can achieve a prediction accuracy of more than 90% in most cases. We also computed the importance of different attributes in the dataset using information gain. This can help the policymakers focus on only certain attributes to study this disease rather than focusing on multiple random factors that may not be very informative to patient outcomes.Graphical abstract

Clustering a database of optically absorbing organic molecules via a hierarchical fingerprint scheme that categorizes similar functional molecular fragments.

A measure of chemical similarity is only useful if it implies similarity in some relevant property space. Typically, similarity calculations operate by assigning each molecule a chemical fingerprint: a fixed-length vector of bits where the on-bits signify the presence of a certain feature. Common fingerprinting schemes, such as extended-connectivity fingerprints, are by definition general and fail to capture much of the domain-specific theory that underpins similarity in a specific domain. In this work, a hierarchical fingerprinting scheme is developed that is bespoke to a database of ∼4500 organic molecules and their cognate optical absorption spectral properties. Our fingerprinting scheme incorporates molecular fragmentation and domain-specific chemical intuition into an algorithm that categorizes each fragment as being one of a core chemical group, a substituent, or a bridge. The algorithm is applied to every molecule in the database to generate a pool of chemically relevant fragments that are labeled according to their structural category. The fingerprint of each molecule is then composed of a nested Python dictionary specifying the unique identifiers of its constituent fragment entities and the structural links between them to give a hierarchical molecular encoding scheme. Four case studies show the application of our fingerprinting scheme to the subject database. In each case, the clustered molecules display a host of interesting chemical trends. The application that was used to develop and implement this bespoke fingerprinting scheme, referred to as ChemCluster, also exposes a host of other cheminformatics tools pertaining to this database, a selection of which is demonstrated in this work. The enhanced similarity comparisons afforded by our fingerprinting scheme, as well as the large repository of categorized fragments generated during its development, constitute the first step toward using this database in a data-driven materials discovery workflow.

A Tailings Dam Long-Term Deformation Prediction Method Based on Empirical Mode Decomposition and LSTM Model Combined with Attention Mechanism

Tailings dams are constructed as storage dams for ore waste, serving as industrial waste piles and for drainage. The dam is negatively affected by rainfall, infiltration lines and its own gravity, which can cause its instability to gradually increase, leading to dam deformation. To predict the irregular changes of tailings dam deformation, empirical mode decomposition (EMD) is applied to the deformation data to obtain the trend and periodic components. The attention mechanism is used to assign different weights to the input variables to overcome the limitation that the long short-term memory (LSTM) model can only generate fixed-length vectors. The lagged autocorrelation coefficient is applied to each decomposed subregion to solve the lagging effect of external factors on dam deformation. Finally, the model is used to predict deformation in multiple directions to test the generalization ability. The proposed method can effectively mitigate the problems of gradient disappearance and gradient explosion. The applied results show that, compared with the control model EMD-LSTM, the evaluation indexes RMSE and MAE improve 23.66% and 27.90%, respectively. The method also has a high prediction accuracy in the remaining directions of the tailings dam, which has a wide practical application effect and provides a new idea for tailings dam deformation mechanism research.

Complex graph convolutional network for link prediction in knowledge graphs

Knowledge graph (KG) embedding models map nodes and edges to fixed-length vectors and obtain the similarity of nodes as the output of a scoring function to predict missing links between nodes. KG embedding methods based on graph convolutional networks (GCNs) have recently gained significant attention due to their ability to add information of neighboring nodes into the nodes’ embeddings. However, existing GCNs are primarily based on real-valued embeddings, which have high distortion, particularly when modeling graphs with varying geometric structures. In this paper, we propose complex graph convolutional network (ComplexGCN), a novel extension of the standard GCNs in complex space to combine the expressiveness of complex geometry with GCNs for improving the representation quality of KG components. The proposed ComplexGCN comprises a set of complex graph convolutional layers and a complex scoring function based on PARATUCK2 decomposition: the former includes information of neighboring nodes into the nodes’ embeddings, while the latter leverages these embeddings to predict new links between nodes. The proposed model demonstrates enhanced performance compared to existing methods on the two recent standard link prediction datasets.

A novel multi-innovation gradient support vector machine regression method

For the regression problem of support vector machine, the solution processes of the most existing methods use offline datasets, which cannot be realized online. For this problem, this paper presents a new online approach to identify these unknown parameters contained in the support vector machine. A new cost function is constructed by substituting the error term into the standard cost function, which is different from the standard support vector machine, and the gradient descent approach is then used to minimize the newly created loss function, thus proposing a stochastic gradient support vector machine algorithm to estimate the unknown parameters based on the recursive identification methods. Furthermore, to advance the property of the stochastic gradient support vector machine algorithm, a moving data window is used to widen the scalar information into a fixed-length innovation vector, thereby increasing the amount of information used in the parameter estimation based on the multi-innovation identification theory. In addition, the forgetting factor is brought into the proposed algorithms, and the corresponding forgetting factor recursive algorithms are derived. These methods are recursive identification methods, which may be implemented online and are more efficient in terms of computing. Finally, utilizing the MatLab platform, the validity and usefulness of the explored methodologies are proven using several numerical simulation examples.

PWM2Vec: An Efficient Embedding Approach for Viral Host Specification from Coronavirus Spike Sequences.

Simple SummaryThe family of coronaviruses comprises a diverse set of strains and variants which cause diseases from the common cold to COVID-19. Moreover, they infect a wide array of hosts from bats, camels, birds, to humans. Studying coronaviruses through the lens of host specificity provides a unique perspective to understanding the evolution, diversity and dynamics of this family. In particular, this can reveal groups of different hosts infected by similar strains, giving clues on strains which were more likely to have evolved to jump from one host to another. In this work, we frame host specificity as a classification task, in designing a very compact numerical representation of the spike sequences of different coronaviruses. Based on this numerical representation, classification methods are able to detect the target host with high accuracy. Such an approach can used to efficiently scale to large volumes of sequences, in order to unveil trends in the host specificity of different coronavirus strains.The study of host specificity has important connections to the question about the origin of SARS-CoV-2 in humans which led to the COVID-19 pandemic—an important open question. There are speculations that bats are a possible origin. Likewise, there are many closely related (corona)viruses, such as SARS, which was found to be transmitted through civets. The study of the different hosts which can be potential carriers and transmitters of deadly viruses to humans is crucial to understanding, mitigating, and preventing current and future pandemics. In coronaviruses, the surface (S) protein, or spike protein, is important in determining host specificity, since it is the point of contact between the virus and the host cell membrane. In this paper, we classify the hosts of over five thousand coronaviruses from their spike protein sequences, segregating them into clusters of distinct hosts among birds, bats, camels, swine, humans, and weasels, to name a few. We propose a feature embedding based on the well-known position weight matrix (PWM), which we call PWM2Vec, and we use it to generate feature vectors from the spike protein sequences of these coronaviruses. While our embedding is inspired by the success of PWMs in biological applications, such as determining protein function and identifying transcription factor binding sites, we are the first (to the best of our knowledge) to use PWMs from viral sequences to generate fixed-length feature vector representations, and use them in the context of host classification. The results on real world data show that when using PWM2Vec, machine learning classifiers are able to perform comparably to the baseline models in terms of predictive performance and runtime—in some cases, the performance is better. We also measure the importance of different amino acids using information gain to show the amino acids which are important for predicting the host of a given coronavirus. Finally, we perform some statistical analyses on these results to show that our embedding is more compact than the embeddings of the baseline models.

Compact feature hashing for machine learning based malware detection

Machine learning can detect variant malware files that can evade signature-based detection. Feature hashing is used to convert features into a fixed-length vector. In this paper, we study the appropriate vector size for feature hashing for a large dataset of malware files. Through exhaustive experiments on more than 280,000 real malware and benign files, we find for the first time that the default vector size of current feature hashing practices is unnecessarily large. We experimentally explore the appropriate vector size, which not only reduces memory space by 70% but also increases the detection accuracy, compared with the state-of-the-art scheme.

Spherical harmonic shape descriptors of nodal force demands for quantifying spatial truss connection complexity

The connections of a spatial truss structure play a critical role in the safe and efficient transfer of axial forces between members. For discrete connections, they can also improve construction efficiency by acting as registration devices that lock members in precise orientations. As more geometrically complex spatial trusses are enabled by computational workflows and the demand for material-efficient spanning systems, there is a need to understand the effects of global form on the demands at the connections. For large-scale structures with irregular geometry, customizing individual nodes to meet exact member orientations and force demands may be infeasible; conversely, standardizing all connections results in oversized nodes and a compromise in registration potential. We propose a method for quantifying the complexity of spatial truss designs by the variation in nodal force demands. By representing nodal forces as a geometric object, we leverage the spherical harmonic shape descriptor, developed for applications in computational geometry, to characterize each node by a rotation and translation-invariant fixed-length vector. We define a complexity score for spatial truss design by the variance in the positions of the feature vectors in higher-dimensional space, providing an additional performance metric during early stage design exploration. We then develop a pathway towards reducing complexity by clustering nodes with respect to their feature vectors to reduce the number of unique connectors for design while minimizing the effects of mass standardization.

Interactive Visual Pattern Search on Graph Data via Graph Representation Learning.

Graphs are a ubiquitous data structure to model processes and relations in a wide range of domains. Examples include control-flow graphs in programs and semantic scene graphs in images. Identifying subgraph patterns in graphs is an important approach to understand their structural properties. We propose a visual analytics system GraphQ to support human-in-the-loop, example-based, subgraph pattern search in a database containing many individual graphs. To support fast, interactive queries, we use graph neural networks (GNNs) to encode a graph as fixed-length latent vector representation, and perform subgraph matching in the latent space. Due to the complexity of the problem, it is still difficult to obtain accurate one-to-one node correspondences in the matching results that are crucial for visualization and interpretation. We, therefore, propose a novel GNN for node-alignment called NeuroAlign, to facilitate easy validation and interpretation of the query results. GraphQ provides a visual query interface with a query editor and a multi-scale visualization of the results, as well as a user feedback mechanism for refining the results with additional constraints. We demonstrate GraphQ through two example usage scenarios: analyzing reusable subroutines in program workflows and semantic scene graph search in images. Quantitative experiments show that NeuroAlign achieves 19%-29% improvement in node-alignment accuracy compared to baseline GNN and provides up to 100× speedup compared to combinatorial algorithms. Our qualitative study with domain experts confirms the effectiveness for both usage scenarios.

Water as a Lévy Rotor.

A probability density function describing the angular evolution of a fixed-length atom-atom vector as a Lévy rotor is derived containing just two dynamical parameters: the Lévy parameter α and a rotational time constant τ. A Lévy parameter α<2 signals anomalous (non-Brownian) motion. Molecular dynamics simulation of water at 298K validates the probability density function for the intramolecular ^{1}H─^{1}H dynamics. The rotational dynamics of water is found to be approximately Brownian at subpicosecond time intervals, becomes increasingly anomalous at longer time intervals due to hydrogen-bond breaking and reforming, before becoming indistinguishable from Brownian dynamics beyond about 25ps. The Lévy rotor model is used to estimate the intramolecular contribution to the longitudinal nuclear-magnetic-resonance (NMR) relaxation rate R_{1,intra}. It is found that R_{1,intra} contributes 65%±7% to the overall relaxation rate of water at room temperature.

Scale-invariant histogram of oriented gradients: novel approach for pedestrian detection in multiresolution image dataset

This paper proposes a scale-invariant histogram of oriented gradients (SI-HOG) for pedestrian detection. Most of the algorithms for pedestrian detection use the HOG as the basic feature and combine other features with the HOG to form the feature set, which is usually applied with a support vector machine (SVM). Hence, the HOG feature is the most efficient and fundamental feature for pedestrian detection. However, the HOG feature produces feature vectors of different lengths for different image resolutions; thus, the feature vectors are incomparable for the SVM. The proposed method forms a scale-space pyramid wherein the histogram bin is calculated. Thus, the gradient information from all the scales is encapsulated in a single fixed-length feature vector. The proposed method is also combined with color and texture features. The proposed approach is tested on three established benchmark pedestrian datasets: INRIA, NICTA, and Daimler. An improvement of ≥4.5% in the miss rate is achieved for all the three datasets considered. We also show that the SI-HOG can be applied to multiresolution datasets for which the HOG feature cannot be applied. Additionally, the MapReduce model is used to obtain the same outcome. The results indicate that the proposed approach outperforms the pedestrian-detection methods considered in this work.

Robust Representation and Efficient Feature Selection Allows for Effective Clustering of SARS-CoV-2 Variants

The widespread availability of large amounts of genomic data on the SARS-CoV-2 virus, as a result of the COVID-19 pandemic, has created an opportunity for researchers to analyze the disease at a level of detail, unlike any virus before it. On the one hand, this will help biologists, policymakers, and other authorities to make timely and appropriate decisions to control the spread of the coronavirus. On the other hand, such studies will help to more effectively deal with any possible future pandemic. Since the SARS-CoV-2 virus contains different variants, each of them having different mutations, performing any analysis on such data becomes a difficult task, given the size of the data. It is well known that much of the variation in the SARS-CoV-2 genome happens disproportionately in the spike region of the genome sequence—the relatively short region which codes for the spike protein(s). In this paper, we propose a robust feature-vector representation of biological sequences that, when combined with the appropriate feature selection method, allows different downstream clustering approaches to perform well on a variety of different measures. We use such proposed approach with an array of clustering techniques to cluster spike protein sequences in order to study the behavior of different known variants that are increasing at a very high rate throughout the world. We use a k-mers based approach first to generate a fixed-length feature vector representation of the spike sequences. We then show that we can efficiently and effectively cluster the spike sequences based on the different variants with the appropriate feature selection. Using a publicly available set of SARS-CoV-2 spike sequences, we perform clustering of these sequences using both hard and soft clustering methods and show that, with our feature selection methods, we can achieve higher F1 scores for the clusters and also better clustering quality metrics compared to baselines.

Statistical and Visual Analysis of Audio, Text, and Image Features for Multi-Modal Music Genre Recognition.

We present a multi-modal genre recognition framework that considers the modalities audio, text, and image by features extracted from audio signals, album cover images, and lyrics of music tracks. In contrast to pure learning of features by a neural network as done in the related work, handcrafted features designed for a respective modality are also integrated, allowing for higher interpretability of created models and further theoretical analysis of the impact of individual features on genre prediction. Genre recognition is performed by binary classification of a music track with respect to each genre based on combinations of elementary features. For feature combination a two-level technique is used, which combines aggregation into fixed-length feature vectors with confidence-based fusion of classification results. Extensive experiments have been conducted for three classifier models (Naïve Bayes, Support Vector Machine, and Random Forest) and numerous feature combinations. The results are presented visually, with data reduction for improved perceptibility achieved by multi-objective analysis and restriction to non-dominated data. Feature- and classifier-related hypotheses are formulated based on the data, and their statistical significance is formally analyzed. The statistical analysis shows that the combination of two modalities almost always leads to a significant increase of performance and the combination of three modalities in several cases.

Deep face fuzzy vault: Implementation and performance

Biometric technologies, especially face recognition, have become an essential part of identity management systems worldwide. In deployments of biometrics, secure storage of biometric information is necessary in order to protect the users’ privacy. In this context, biometric cryptosystems are designed to meet key requirements of biometric information protection enabling a privacy-preserving storage and comparison of biometric data, e.g. feature vectors extracted from facial images. Until now, biometric cryptosystems have hardly been applied to state-of-the-art biometric recognition systems utilizing deep convolutional neural networks.This work investigates the application of a well-known biometric cryptosystem, i.e. the improved fuzzy vault scheme, to facial feature vectors extracted through deep convolutional neural networks. To this end, a feature transformation method is introduced which maps fixed-length real-valued deep feature vectors to integer-valued feature sets. As part of said feature transformation, a detailed analysis of different feature quantisation and binarisation techniques is conducted. At key binding, obtained feature sets are locked in an unlinkable improved fuzzy vault. For key retrieval, the efficiency of different polynomial reconstruction techniques is investigated. The proposed feature transformation method and template protection scheme are agnostic of the biometric characteristic and, thus, can be applied to virtually any biometric features computed by a deep neural network. In experiments, an unlinkable improved deep face fuzzy vault-based template protection scheme is constructed employing features extracted with a state-of-the-art deep convolutional neural network trained with the additive angular margin loss (ArcFace). For the best configuration, a false non-match rate below 1% at a false match rate of 0.01%, is achieved in cross-database experiments on the FERET and FRGCv2 face databases. On average, a security level of up to approximately 28 bits is obtained. This work presents an effective face-based fuzzy vault scheme providing privacy protection of facial reference data as well as digital key derivation from face.