Search

There is an accumulating evidence that distracted driving is a leading cause of vehicle crashes and accidents. In order to support safe driving, numerous methods of detecting distraction have been proposed, which are empirically focused on certain driving contexts and gaze behaviour. This paper aims at illustrating a method for the non-intrusive and real-time detection of visual distraction based on vehicle dynamics data and environmental data, without using eye-tracker information. Experiments are carried out in the context of the automotive domain of the European project Holides, which addresses development and qualification of adaptive cooperative human–machine systems, and is co-funded by ARTEMIS Joint Undertaking and Italian University, Educational and Research Department. The collected data are analysed by a single-layer feedforward neural network trained through pseudo-inversion methods, characterized by direct determination of output weights given randomly set input weights and biases. One main feature of our work is the convenient setting of input weights by the so-called sparse random projections: the presence of a great number of null elements in the involved matrices makes especially parsimonious the use at run time of the trained network. Moreover, we use a genetic approach to better explore the input weights network space. The obtained results show better performance with respect to classical pseudo-inversion methods and effective and parsimonious use of memory resources.

In the last decade the use of artificial neural networks (ANNs) in many fields like image processing or speech recognition has become a common practice because of their effectiveness to solve complex tasks. However, in such a rush, very little attention has been paid to security aspects. In this work we explore the possibility to embed a watermark into the ANN parameters. We exploit model redundancy and adaptation capacity to lock a subset of its parameters to carry the watermark sequence. The watermark can be extracted in a simple way to claim copyright on models but can be very easily attacked with model fine-tuning. To tackle this culprit we devise a novel watermark aware training strategy. We aim at delving into the loss landscape to find an optimal configuration of the parameters such that we are robust to fine-tuning attacks towards the watermarked parameters. Our experimental results on classical ANN models trained on well-known MNIST and CIFAR-10 datasets show that the proposed approach makes the embedded watermark robust to fine-tuning and compression attacks.

We present an algorithm for fragile watermarking of color, or multi-channel, images either in uncompressed format, in lossless compressed format, or in compressed format with locally compressed units (like JPEG). The watermark is embedded into the Karhunen-Loeve transform (KLT) coefficients of the host image, and the security of the method is based on the secrecy of the KLT basis computed from a corresponding secret key image. The watermark bits may be embedded with various methods, in particular the use of syndrome coding has proven flexible and effective. Genetic Algorithms (GAs) are used to find the optimum pixel modification leading to the watermark embedding. The resulting watermarked images have shown to have both a high objective quality (in terms of PSNR and SSIM) and a high subjective quality (tested by a group of observers). At the same time, the watermark extraction is very sensitive to small (?1 intensity level for single pixel) modifications, ensuring image authentication with very high probability. We propose an image fragile watermarking framework in the KLT transformed domain.The modular structure of the proposed algorithm makes it flexible and easily extendable.An optimization algorithm saves the use of inverse transformation.It works and produces both lossless and lossy compressed color images.The quality of the resulting images is very high.

Machine learning (ML) components are increasingly adopted in many automated systems. Their ability to learn and work with novel input/incomplete knowledge and their generalization capabilities make them highly desirable solutions for complex problems. This has motivated the inclusion of ML techniques/components in products for many industrial domains including automotive systems. Such systems are safety-critical systems since their failure may cause death or injury to humans. Therefore, their safety must be ensured before they are used in their operational environment. However, existing safety standards and Verification and Validation (V&V) techniques do not properly address the special characteristics of ML-based components such as non-determinism, non-transparency, instability. This position paper presents the authors' view on the safety of automotive systems incorporating ML-based components, and it is intended to motivate and sketch a research agenda for extending a safety standard, namely ISO 26262, to address challenges posed by incorporating ML-based components in automotive systems.

In the context of an increasing dependence on multimedia contents, data hiding techniques, such as watermarking and steganography, are becoming more and more important. Due to the complementary nature of their general requirements, i.e., imperceptibility, robustness, security and capacity, many data hiding schemes endeavour to find the optimal performances by applying various approaches inspired from nature. In this paper, we provide a review and analysis of the main computational intelligence approaches, including Artificial Neural Networks (ANNs) and Fuzzy Sets (FSs), which are employed in information hiding. Furthermore, with the aid of the recent state of the art, we discuss the main challenges to be addressed and future directions of research.

Fragile watermarking has attracted a lot of attention in the last decade. An interesting approach, presented in 2011 by Lin et al., results in very high quality of the watermarked images. However, after a thorough examination of the paper, a few improvements are proposed in our revised version of the algorithm in order to overcome some shortcomings. In particular, changes to the pseudocode and modifications to deal with pixel saturation are suggested, along with a way to improve the scheme security. Finally, a deeper analysis of the security is presented.

In recent years, the KDD process has been advocated to be an iterative and interactive process. It is seldom the case that a user is able to answer immediately all his questions on date with a single query. On the contrary, the work-flow of the typical user consists of several steps in which he/she iteratively refines the extracted knowledge by inspecting previous results and posing new queries. Given this view of the KDD process, in order to reduce the computational effort, it becomes crucial to have KDD systems that are able to exploit past results. This is especially true in environments in which the system knowledge base is the result of many discoveries on data made separately by the collaborative effort of different users. In this paper, we consider the problem of mining frequent association rules from database relations. We first model a general, constraint-based, mining language for this task. Then, we propose an algorithm that answers such queries reusing past results. In particular, this solution is effective for a new class of constraints, called context dependent, which are more difficult than the traditionally studied item dependent constraints. Nevertheless, we show that some typical queries of important application domains, such as market stock trading, analysis of web log, and gene microarrays in bioinformatics, have context-dependent constraints. We show with a set of experiments in these application domains that the proposed solution with an incremental approach is both effective and viable.

The proliferation of Infrared technology and imaging systems enables a different perspective to tackle many computer vision problems in defense and security applications. Infrared images are widely used by the law enforcement, Homeland Security and military organizations to achieve a significant advantage or situational awareness, and thus is vital to protect these data against malicious attacks. Concurrently, sophisticated malware are developed which are able to disrupt the security and integrity of these digital media. For instance, illegal distribution and manipulation are possible malicious attacks to the digital objects. In this paper we explore the use of a new layer of defense for the integrity of the infrared images through the aid of information hiding techniques such as watermarking. In this context, we analyze the efficiency of several optimal decoding schemes for the watermark inserted into the Singular Value Decomposition (SVD) domain of the IR images using an additive spread spectrum (SS) embedding framework. In order to use the singular values (SVs) of the IR images with the SS embedding we adopt several restrictions that ensure that the values of the SVs will maintain their statistics. For both the optimal maximum likelihood decoder and sub-optimal decoders we assume that the PDF of SVs can be modeled by the Weibull distribution. Furthermore, we investigate the challenges involved in protecting and assuring the integrity of IR images such as data complexity and the error probability behavior, i.e., the probability of detection and the probability of false detection, for the applied optimal decoders. By taking into account the efficiency and the necessary auxiliary information for decoding the watermark, we discuss the suitable decoder for various operating situations. Experimental results are carried out on a large dataset of IR images to show the imperceptibility and efficiency of the proposed scheme against various attack scenarios.

This paper describes a module that extracts rules or frequent patterns through data mining from a large database fed by targets detected by a Mission System installed on an unmanned airborne platform and the associated ground station to discover anomalies in local traffic. It has been demonstrated that the module is able to detect all tracks or targets present in the ground truth and also the paths followed by each tracks. Traffic anomalies can be detected by observing differences in extracted rules in reference missions compared to the current mission. The module will significantly reduce the operator workload as it can operate autonomously‥

In this paper we show that a recently proposed fragile watermarking scheme by Rawat et al. does not detect and localize tampering, therefore cannot be used for authentication applications. The problem lies in that the scheme embeds an authentication code into the LSBs of pixels without taking into consideration the image content. To overcome this issue the authentication should be combined with the first seven bits of the image pixels, and in this paper a revision in this sense is proposed.

Genetic Algorithms (GAs) are known to be valuable tools for optimization purposes. In general, GAs can find good solutions by setting their configuration parameters, such as mutation and crossover rates, population size, etc., to standard (i.e., widely used) values. In some application domains, changing the values of these parameters does not improve the quality of the solution, but might influence the ability of the algorithm to find such solution. In other application domains, fine tuning these parameters could result into a significant improvement of the solution quality. In this paper we present an experimental study aimed at finding how fine tuning the parameters of a GA used for the insertion of a fragile watermark into a bitmap image influences the quality of the resulting digital object. However, when proposing a GA based new tool to non-expert users, selecting the best parameter setting is not an easy task. Therefore, we will suggest how to automatically set the GA parameters in order to meet the quality and/or running time performances requested by the user.

There is accumulating evidence that driver distraction is a leading cause of vehicle crashes and incidents. In particular, increased use of so-called in-vehicle information systems (IVIS) and partially autonomous driving assistance systems (PADAS) have raised important and growing safety concerns. Thus, detecting the driver's state is of paramount importance, to adapt IVIS and PADAS accordingly, therefore avoiding or mitigating their possible negative effects. The purpose of this paper is to show a method for the nonintrusive and real-time detection of visual distraction, using vehicle dynamics data and without using the eye-tracker data as inputs to classifiers. Specifically, we present and compare different models that are based on well-known machine learning (ML) methods. Data for training the models were collected using a static driving simulator, with real human subjects performing a specific secondary task [i.e., a surrogate visual research task (SURT)] while driving. Different training methods, model characteristics, and feature selection criteria have been compared. Based on our results, using a support vector machine (SVM) has outperformed all the other ML methods, providing the highest classification rate for most of the subjects. Potential applications of this paper include the design of an adaptive IVIS and of a “smarter” PADAS.

Discovery biological organisation of the cell in modules network is a challenging task. Currently, approaches based on a controlled vocabulary, as Gene Ontology, to identify the function similarity among a pair of genes have developed. We present RGO: a rewriting of the GO aiming at obtaining a more compact and informative ontology, leading to closer biological regulated GO terms. The RGO will help researcher to easily identify the genes belonging to the same network module without the need of additional data.

The problem of learning a classifier from examples is a fundamental task in Machine Learnig and is nowadays actively studied. Most approaches follow a Discrimination based paradigm, where the aim is to find the best way to separate the examples of one class from those of other classes. A less popular approach follows a Characterization based paradigm, where the aim is to build a description (model) of every class from examples, and use this model to recognizes instances of that class. This paper focuses on the specific task of classifying and tagging symbolic sequences, by introducing a Characterization approach, based on Structured Hidden Markov Models, and compares its perfomances against a widely used discriminative approach, i.e. kernel machines. This task is particulary relevant to several applications in many fields, such as molecular biology, web log analysis, network traffic analysis, user profiling, and so on. In order to assess the validity of the proposed approach, an artificial benchmark has been designed in such a way that the regularities to be discovered are well known and allow for a controlled evaluation of the real capabilities the learning algorithms investigated. The obtained results allow to point out the major advantages and weaknesses of the investigated approaches in the specific classification task addressed.

This paper presents a new approach to learning diagnostic knowledge, based on the use of a causal model of the domain and abductive reasoning. Justifications supplied by the casual model serve both to focus the search during one-step learning and to localize failures and propose changes during knowledge refinement.

The huge volume of gene expression data produced by microarrays and other high-throughput techniques has encouraged the development of new computational techniques to evaluate the data and to formulate new biological hypotheses. To this purpose, co-clustering techniques are widely used: these identify groups of genes that show similar activity patterns under a specific subset of the experimental conditions by measuring the similarity in expression within these groups. However, in many applications, distance metrics based only on expression levels fail in capturing biologically meaningful clusters. We propose a methodology in which a standard expression-based co-clustering algorithm is enhanced by sets of constraints which take into account the similarity/dissimilarity (inferred by the Gene Ontology, GO) between pairs of genes. Our approach minimizes the intervention of the analyst within the co-clustering process. It provides meaningful co-clusters whose discovery and interpretation is increased by embedding GO annotations.

The goal of this paper is to apply some of the best-known machine learning techniques to a practical problem in the automotive field: the identification and classification of the user's intentions in performing specific driving maneuvers. Data have been collected by a static driving simulator. These models are then analyzed and compared, in order to select the best car-following maneuver classifier.

Distraction during driving task is one of the most serious problems affecting traffic safety, being one of the main causes of accidents. Therefore, a method to diagnose and evaluate Distraction appears to be of paramount importance to study and implement efficient counter-measures. This research aims at illustrating our approach in diagnosis of Distraction status, comparing some of the widely used data-mining techniques; in particular, Fuzzy Logic (with Adaptive-Network-based Fuzzy Inference System) and Artificial Neural Networks. The results are compared to select which method gives the best performances.

In this paper we present a stepwise method for the analysis of musical sequences. The starting point is either a MIDI file or the score of a piece of music. The result is a set of likely themes and motifs. The method relies on a pitch intervals representation of music and an event discovery system that extracts significant and repeated patterns from sequences. We report and discuss the results of a preliminary experimentation, and outline future enhancements.

In recent years, the KDD process has been advocated to be an iterative and interactive process. It is seldom the case that a user is able to answer immediately with a single query all his questions on data. On the contrary, the workflow of the typical user consists in several steps, in which he/she iteratively refines the extracted knowledge by inspecting previous results and posing new queries. Given this view of the KDD process, it becomes crucial to have KDD systems that are able to exploit past results thus minimizing computational effort. This is expecially true in environments in which the system knowledge base is the result of many discoveries on data made separately by the collaborative effort of different users. In this paper, we consider the problem of mining frequent association rules from database relations. We model a general, constraint-based, mining language for this task and study its properties w.r.t. the problem of re-using past results. In particular, we individuate two class of query constraints, namely “item dependent” and “context dependent” ones, and show that the latter are more difficult than the former ones. Then, we propose two newly developed algorithms which allow the exploitation of past results in the two cases. Finally, we show that the approach is both effective and viable by experimenting on some datasets.

A number of techniques have been developed in order to address issues such as genome, trascriptome and proteome analysis. However, a time and cost effective technique for interactome analysis is still lacking. Lots of methods for the predicion of protein - protein interacions have been developed: some of them are based on high quality alignment of sequences, others are based on the tridimensional features of proteins, but they all bear strong limitations that make impossible their large scale application. Recently, an SVM-based machine learning approach has been used to address this topic. Although the method was able to correctly classify 80% of the test samples, it was not applied to the prediction of yet unknown interactions. In this work, we address this topic and show that an optimized, SVM-based machine learning approach trained with combinations of shuffled sequences as examples of lack of interaction is unable to make large scale predictions of interaction. © Springer-Verlag Berlin Heidelberg 2003.

Recently inductive databases (IDBs) have been proposed to afford the problem of knowledge discovery from huge databases. With an IDB the user/analyst performs a set of very different operations on data using a special-purpose language, powerful enough to perform all the required manipulations, such as data preprocessing, pattern discovery and pattern post-processing. In this paper we present a comparison between query languages (MSQL, DMQL and MINE RULE) that have been proposed for association rules extraction in the last years and discuss their common features and differences. We present them using a set of examples, taken from the real practice of data mining. This allows us to define the language design guidelines, with particular attention to the open issues on IDBs.

Boosting is a powerful and thoroughly investigated learning technique that improves the accuracy of any given learning algorithm by weighting training examples and hypotheses. Several authors contributed to the general boosting learning framework with theoretical and experimental results, mainly in the propositional learning framework. In a previous paper, we investigated the applicability of Freund and Schapire's AdaBoost.M1 algorithm to a first order logic weak learner. In this paper, we extend the weak learner in order to directly deal with weighted instances and compare two ways to apply boosting to such a weak learner: resampling instances at each round and using weighted instances.

This paper presents the experimental evaluation of a neural network architecture that can manage structured data and refine knowledge bases expressed in a first order logic language.

RTL is an algorithm designed to learn any number of simple, mutually dependent relations, producing recursive programs that are stratified in the sense given by Apt. In this paper, we present a revised algorithm and its implementation based on previous theoretical works that establish properties and limits of the learning framework. The algorithm is described both in abstract form and through an example. Emphasis is put on the way RTL uses induction and domain knowledge to guide the search towards specific kinds of hypothesis. The algorithm has been tested on three different domains obtaining encouraging results, as reported in the discussion. Finally, it is shown experimentally that the control strategy realized is somewhat independent of the order in which concepts are learned.

In the last decade, many efforts have been devoted to the exploration of techniques for learning and refining first order theories, as the necessary step for applying machine learning methodologies to real world applications. In this paper, we present a new approach to the integration of inductive and deductive learning techniques that seems to overcome some of the limitations of existing learning systems without imposing strong constraints or biases on both the representation language and the search space. In particular, a new search structure that enables the system to learn a structured knowledge base is proposed. Moreover, the learning system described in the paper can be used both to learn new knowledge from scratch and to refine an existing background theory.

Pattern classification is considered a major task both in Artificial Intelligence and in Pattern Recognition. Artificial Intelligence developed symbolic methods for learning classification rules which are are effective in order to acquire the structure of the knowledge, whereas Pattern Recognition developed methods effective for learning numerical features such as weights and thresholds. However, both structural and numeric knowledge is involved in classifying complex patterns. This paper presents a method for dealing with numeric features in a learning framework based on a first order logic. The method has been implemented and works in two steps. In the first step, tentative numeric features are learned together with symbolic features, whereas in the second one the numeric knowledge is refined using a standard genetic algorithm.

In recent years, researchers have begun to study inductive databases, a new generation of databases for leveraging decision support applications. In this context, the user interacts with the DBMS using advanced, constraint-based languages for data mining where constraints have been specifically introduced to increase the relevance of the results and, at the same time, to reduce its volume. In this paper we study the problem of mining frequent itemsets using an inductive database 1 . We propose a technique for query answering which consists in rewriting the query in terms of union and intersection of the result sets of other queries, previously executed and materialized. Unfortunately, the exploitation of past queries is not always applicable. We then present sufficient conditions for the optimization to apply and show that these conditions are strictly connected with the presence of functional dependencies between the attributes involved in the queries. We show some experiments on an initial prototype of an optimizer which demonstrates that this approach to query answering is not only viable but in many practical cases absolutely necessary since it reduces drastically the execution time.