Intelligence artificielle et données

This document demonstrates the implementation of a Spark batch processing example using distributed computing paradigms for handling large datasets. It provides detailed instructions for setting up the environment and running the batch processes efficiently. Key methodologies include data ingestion, transformation, and storage using the Spark framework. The findings emphasize the scalability and performance benefits of Spark for batch processing tasks in real-world scenarios.

The document introduces the concept of Big Data by explaining its origins, characteristics, and challenges it imposes on traditional data management systems. It discusses the core 'Vs' of Big Data (Volume, Velocity, Variety) and their consequences on system architecture, such as the need for horizontal scalability, continuous availability, and flexibility. The text also covers the CAP theorem and its implications for distributed Big Data systems, highlighting choices between consistency, availability, and partition tolerance. Solutions like distributed and scalable infrastructures are neces...

Decision tree learning is a predictive modeling approach commonly used in statistics, data mining, and machine learning. It employs a tree structure where observations determine branches and conclusions determine leaf values. Key strengths include intelligible knowledge validation, automatic variable selection, and robust performance against outliers. However, weaknesses include instability with small datasets, difficulty in identifying variable interactions, and masking of selected variable importance.

The document explores decision trees with the Gini index using a banking example to predict loan repayment success. The methodology includes stepwise tree construction and conversion into rule bases, highlighting strengths like variable selection and non-parametric nature. Weaknesses like instability on small datasets and difficulty capturing variable interactions are also discussed. The analysis emphasizes practical benefits such as robustness to outliers and medium dataset processing speed.

The document focuses on analyzing data related to socio-professional groups and their consumption patterns for various food items. Principal Component Analysis (PCA) is utilized to identify underlying patterns and correlations among variables. The correlation matrix reveals significant associations, with 'Autre pain' and 'Raisin de table' exhibiting the highest correlation (0.959), while 'Autre pain' and 'Jus ordinaire' show the least. Total eigenvalue analysis suggests two primary factorial axes contributing to 88.59% of variability. The study interprets these axes, revealing distinct grou...

This paper introduces the Universal Trigger (UniTrigger), a powerful adversarial textual attack method that can significantly impact neural network prediction accuracy. The authors propose DARCY, a honeypot-based defense framework designed to detect these attacks by injecting multiple trapdoors into the model. Comprehensive experiments show that DARCY achieves high detection rates with minimal impact on prediction accuracy for clean inputs.

This document discusses the concepts of adversarial robustness in machine learning, highlighting the brittleness of ML models and the security implications of adversarial examples. It examines the limitations of supervised learning frameworks, the challenges posed by data poisoning, and the importance of understanding failure modes in ML systems.

This paper explores the phenomenon of adversarial examples in machine learning, attributing their existence to non-robust features that are brittle and perplexing to humans. The authors present a theoretical framework illustrating the prevalence of these features across standard datasets and demonstrate a misalignment between human notions of robustness and data geometry. The findings suggest a need to reevaluate adversarial vulnerability as a consequence of reliance on well-generalizing features.

This research investigates the effectiveness of combining multiple weak defenses against adversarial examples in neural networks. The authors demonstrate that ensembles of weak defenses do not provide a strong defense, as adaptive adversaries can successfully attack the combined defenses with low distortion. The findings imply that simply aggregating ineffective defenses is insufficient for robust protection against threats in security-critical applications.

This paper discusses the vulnerabilities of deep learning systems to adversarial attacks, which can lead to misclassification of outputs despite being imperceptible to humans. It reviews various types of adversarial attacks based on different threat models and evaluates the efficiency and challenges of existing countermeasures. The authors emphasize the need for robust deep learning algorithms to combat these adversarial threats.

The paper explores a novel approach to detecting adversarial attacks on neural networks by injecting artificial weaknesses, termed 'trapdoors', into trained models. These trapdoors are designed to attract adversarial perturbations, allowing the model owner to recognize attacks through distinctive neuron activation signatures. The authors demonstrate high detection success rates across various classification domains with minimal impact on normal inputs.

This document explores the reasons why classifiers in high dimensions are vulnerable to adversarial perturbations. The authors argue that the vulnerability is likely due to computational constraints rather than information theoretic limitations. They provide a binary classification task that illustrates the exponential separation between classical learning and robust learning in the statistical query model.

This paper proposes a novel Feature Manipulation Defense mechanism to detect and purify adversarial examples in machine learning. By using a combo-variational autoencoder to manipulate semantic features, the method demonstrates high detection and purification accuracy across multiple datasets. The effectiveness of this approach lies in its ability to resist misclassification due to non-significant changes in input features.