Contributed Session 5: Advanced Methodology

Deep neural networks (DNNs) have driven advancements across various domains such as computer vision, natural language processing, and autonomous systems. Despite their success, DNNs are increasingly vulnerable to adversarial attacks, particularly Trojan attacks. These attacks manipulate networks during training to behave maliciously in response to specific triggers while performing normally on clean data. Detecting Trojaned networks is challenging due to subtle modifications and limited access to data, making traditional detection methods reliant on black-box approaches impractical. Topological data analysis (TDA) has emerged as a promising alternative by leveraging Persistent Homology (PH) to identify structural deviations within neural networks. This technique offers a novel way to uncover hidden vulnerabilities by analyzing the intrinsic properties of a model. Previous work has demonstrated the potential of TDA by utilizing summarizing topological features to distinguish between Trojaned and clean models. This work aims to extend topological detection methodologies by proposing a framework that concatenates persistence diagrams across multiple models. By extracting values from these diagrams and converting them into images, each highlighted feature is treated independently for training a machine learning classifier. Our approach aims to enhance the robustness and accuracy of Trojan detection while providing insights into topological differences between clean and Trojaned models.

In this work, we consider observations coming from spatio-temporal point patterns observed in network-constrained regions, such as road networks. We present a nonparametric estimator for probability density functions within a spatio-temporal framework. To this end, we penalize the maximum likelihood estimator using roughness penalties designed to encode smoothness across both spatial and temporal domains. We balance the model’s adherence to the data with the regularity of the estimate by introducing smoothing parameters, which we calibrate appropriately. To solve the estimation problem, we develop a finite element method specifically for networks and employ optimization algorithms, ensuring high accuracy, flexibility, and efficiency, even for large networks.

Effectiveness of oversampling techniques, to address imbalanced classification problems, is usually associated with high imbalance ratios. Nonetheless, it is well recognized that class overlap could play a crucial role in classification performance. This paper investigates how the degree of class overlap influences the effectiveness of oversampling methods, focusing on SMOTE and ROSE. Through a controlled simulation study, we analyze different imbalance ratios under varying levels of class overlap. Our results, although preliminary, suggest that oversampling turns out to be more beneficial when severe class overlap exists. Results also suggest that the choice of oversampling techniques could take advantage by a preliminary measure of class overlap. We discuss potential directions for refining oversampling strategies and propose further research on class overlap quantification.

A discrete probability distribution is proposed, which is defined on the first k positive integers (k >= 3) and derived from the continuous two-parameter Kumaraswamy distribution supported on (0,1). This distribution provides an alternative to existing models, such as the discrete Beta, shifted Beta-Binomial, and CUB distributions, for modeling ordinal data, which are common in many applied disciplines. Key properties of the distribution are investigated, with particular emphasis on the possible shapes of its probability mass function, moments, pseudo-random simulation, and inferential procedures. An analysis of a real data set is included to illustrate the practical application of the proposed model and to evaluate its performance in fitting rating data.