Nicolas Boutry

Transforming gradient-based techniques into interpretable methods

Abstract

The explication of Convolutional Neural Networks (CNN) through xAI techniques often poses challenges in interpretation. The inherent complexity of input features, notably pixels extracted from images, engenders complex correlations. Gradient-based methodologies, exemplified by Integrated Gradients (IG), effectively demonstrate the significance of these features. Nevertheless, the conversion of these explanations into images frequently yields considerable noise. Presently, we introduce GAD (Gradient Artificial Distancing) as a supportive framework for gradient-based techniques. Its primary objective is to accentuate influential regions by establishing distinctions between classes. The essence of GAD is to limit the scope of analysis during visualization and, consequently reduce image noise. Empirical investigations involving occluded images have demonstrated that the identified regions through this methodology indeed play a pivotal role in facilitating class differentiation.

Continue reading

Graph-based spectral analysis for detecting cyber attacks

Abstract

Spectral graph theory delves into graph properties through their spectral signatures. The eigenvalues of a graph’s Laplacian matrix are crucial for grasping its connectivity and overall structural topology. This research capitalizes on the inherent link between graph topology and spectral characteristics to enhance spectral graph analysis applications. In particular, such connectivity information is key to detect low signals that betray the occurrence of cyberattacks. This paper introduces SpectraTW, a novel spectral graph analysis methodology tailored for monitoring anomalies in network traffic. SpectraTW relies on four spectral indicators, Connectedness, Flooding, Wiriness, and Asymmetry, derived from network attributes and topological variations, that are defined and evaluated. This method interprets networks as evolving graphs, leveraging the Laplacian matrix’s spectral insights to detect shifts in network structure over time. The significance of spectral analysis becomes especially pronounced in the medical IoT domains, where the complex web of devices and the critical nature of healthcare data amplify the need for advanced security measures. Spectral analysis’s ability to swiftly pinpoint irregularities and shift in network traffic aligns well with the medical IoT’s requirements for prompt attack detection.

Continue reading

Unsupervised discovery of interpretable visual concepts

Abstract

Providing interpretability of deep-learning models to non-experts, while fundamental for a responsible real-world usage, is challenging. Attribution maps from xAI techniques, such as Integrated Gradients, are a typical example of a visualization technique containing a high level of information, but with difficult interpretation. In this paper, we propose two methods, Maximum Activation Groups Extraction (MAGE) and Multiscale Interpretable Visualization (Ms-IV), to explain the model’s decision, enhancing global interpretability. MAGE finds, for a given CNN, combinations of features which, globally, form a semantic meaning, that we call concepts. We group these similar feature patterns by clustering in concepts, that we visualize through Ms-IV. This last method is inspired by Occlusion and Sensitivity analysis (incorporating causality) and uses a novel metric, called Class-aware Order Correlation (CAOC), to globally evaluate the most important image regions according to the model’s decision space. We compare our approach to xAI methods such as LIME and Integrated Gradients. Experimental results evince the Ms-IV higher localization and faithfulness values. Finally, qualitative evaluation of combined MAGE and Ms-IV demonstrates humans’ ability to agree, based on the visualization, with the decision of clusters’ concepts; and, to detect, among a given set of networks, the existence of bias.

Continue reading

Bridging human concepts and computer vision for explainable face verification

Abstract

With Artificial Intelligence (AI) influencing the decision-making process of sensitive applications such as Face Verification, it is fundamental to ensure the transparency, fairness, and accountability of decisions. Although Explainable Artificial Intelligence (XAI) techniques exist to clarify AI decisions, it is equally important to provide interpretability of these decisions to humans. In this paper, we present an approach to combine computer and human vision to increase the explanation’s interpretability of a face verification algorithm. In particular, we are inspired by the human perceptual process to understand how machines perceive face’s human-semantic areas during face comparison tasks. We use Mediapipe, which provides a segmentation technique that identifies distinct human-semantic facial regions, enabling the machine’s perception analysis. Additionally, we adapted two model-agnostic algorithms to provide human-interpretable insights into the decision-making processes.

Continue reading

Discrete Morse functions and watersheds

Abstract

Any watershed, when defined on a stack on a normal pseudomanifold of dimension $d$, is a pure $(d-1)$-subcomplex that satisfies a drop-of-water principle. In this paper, we introduce Morse stacks, a class of functions that are equivalent to discrete Morse functions. We show that the watershed of a Morse stack on a normal pseudomanifold is uniquely defined, and can be obtained with a linear-time algorithm relying on a sequence of collapses. Last, we prove that such a watershed is the cut of the unique minimum spanning forest, rooted in the minima of the Morse stack, of the facet graph of the pseudomanifold.

Continue reading

Introducing PC $n$-manifolds and $P$-well-composedness in partially ordered sets

Abstract

In discrete topology, discrete surfaces are well-known for their strong topological and regularity properties. Their definition is recursive, and checking if a poset is a discrete surface is tractable. Their applications are numerous: when domain unicoherence is ensured, they lead access to the tree of shapes, and then to filtering in the shape space (shapings); they also lead to Laplacian zero-crossing extraction, to brain tumor segmentation, and many other applications related to mathematical morphology. They have many advantages in digital geometry and digital topology since discrete surfaces do not have any pinches (and then the underlying polyhedron of their geometric realization can be parameterized). However, contrary to topological manifolds known in continuous topology, discrete surfaces do not have any boundary, which is not always realizable in practice (finite hyper-rectangles cannot be discrete surfaces due to their non-empty boundary). For this reason, we propose the three following contributions: (1) we introduce a new definition of boundary, called border, based on the definition of discrete surfaces, and which allows us to delimit any partially ordered set whenever it is not embedded in a greater ambient space, (2) we introduce $P$-well-composedness similar to well-composedness in the sense of Alexandrov but based on borders, (3) we propose new (possibly geometrical) structures called (smooth) $n$-PCM’s which represent almost the same regularity as discrete surfaces and that are tractable thanks to their recursive definition, and (4) we prove several fundamental theorems relative to PCM’s and their relations with discrete surfaces. We deeply believe that these new $n$-dimensional structures are promising for the discrete topology and digital geometry fields.

Continue reading

Structural and spectral analysis of dynamic graphs for attack detection

Abstract

At this time, cyberattacks represent a constant threat. Many approaches exist for detecting suspicious behaviors, but very few of them seem to benefit from the huge potential of mathematical approaches like spectral graph analysis, known to be able to extract topological features of a graph using its Laplacian spectrum. For this reason, we consider our network as a dynamic graph composed of nodes (representing the devices) and of edges (representing the requests), and we compute its Laplacian spectrum across time. An important change of topology inducing an important change in the spectrum, this spectrum seems to be the key to detect threats. Dynamic spectrum-based metrics have been developed for this aim.

Continue reading

Towards attack detection in traffic data based on spectral graph analysis

Abstract

Nowadays, cyberattacks have become a significant concern for individuals, organizations, and governments. These attacks can take many forms, and the consequences can be severe. In order to protect ourselves from these threats, it is essential to employ a range of different strategies and techniques like detection of patterns, classification of system behaviors against previously known attacks, and anomaly detection techniques. This way, we can identify unknown forms of attacks. Few of these existing techniques seem to fully utilize the potential of mathematical approaches such as spectral graph analysis. This domain is made of tools able to extract important topological features of a graph by computing its Laplacian matrix and its corresponding spectrum. This framework can provide valuable insights into the underlying structure of a network, which can be used to detect cyberthreats. Indeed, significant changes in the topology of the graph result in significant changes in the spectrum of the Laplacian matrix. For this reason, we propose here to address this issue by considering the network as a dynamic graph composed of nodes (devices) and edges (requests between devices), to study the evolution of the Laplacian spectrum, and to compute metrics on this evolving spectrum. This way, we should be able to detect suspicious behaviors which may indicate that an attack is occurring.

Continue reading

Gradients intégrés renforcés

Abstract

Les visualisations fournies par les techniques d’Intelligence Artificielle Explicable xAI) pour expliquer les réseaux de neurones convolutionnels (CNN’s) sont parfois difficile á interpréter. La richesse des motifs d’une image qui sont fournis en entrées (les pix l d’une image) entraîne des corrélations complexes entre les classes. Les techniques basées sur les gradients, telles que les gradients intégrés, mettent en évidence l’import nce de ces caractéristiques. Cependant, lorsqu’on les visualise sous forme d’images, on peut e retrouver avec un bruit excessif et donc une difficulté á interpréter les explic tions fournies. Nous proposons la méthode intitulée Gradients Intégrés Renforcés (RI ), une variation des gradients intégrés, qui vise á mettre en évidence les régions nfluentes des images dans la décision des réseaux. Cette méthode vise á réduire la sur ace des zones á analyser lors de la visualisation des résultats, générant ainsi moins e bruit apparent. Des expériences á base d’occlusions démontrent que les régions chois es par notre méthode jouent effectivement un rôle important en terme de classification.

Continue reading

An efficient cascade of U-Net-like convolutional neural networks devoted to brain tumor segmentation

Abstract

Continue reading