State-of-the-art deep neural networks have achieved impressive results on many image classification tasks. However, these same architectures been shown to be unstable small, well sought, perturbations of the images. Despite importance this phenomenon, no effective methods proposed accurately compute robustness state-of-the-art classifiers such large-scale datasets. In paper, we fill gap and propose DeepFool algorithm efficiently that fool networks, thus reliably quantify classifiers....

In applications such as social, energy, transportation, sensor, and neuronal networks, high-dimensional data naturally reside on the vertices of weighted graphs. The emerging field signal processing graphs merges algebraic spectral graph theoretic concepts with computational harmonic analysis to process signals this tutorial overview, we outline main challenges area, discuss different ways define domains, which are analogues classical frequency domain, highlight importance incorporating...

Given a state-of-the-art deep neural network classifier, we show the existence of universal (image-agnostic) and very small perturbation vector that causes natural images to be misclassified with high probability. We propose systematic algorithm for computing perturbations, networks are highly vulnerable such albeit being quasi-imperceptible human eye. further empirically analyze these perturbations show, in particular, they generalize well across networks. The surprising reveals important...

Research in graph signal processing (GSP) aims to develop tools for data defined on irregular domains. In this paper, we first provide an overview of core ideas GSP and their connection conventional digital processing, along with a brief historical perspective highlight how concepts recently developed build top prior research other areas. We then summarize recent advances developing basic tools, including methods sampling, filtering, or learning. Next, review progress several application...

We describe methods for learning dictionaries that are appropriate the representation of given classes signals and multisensor data. further show dimensionality reduction based on dictionary can be extended to address specific tasks such as data analy sis or classification when includes a class separability criteria in objective function. The benefits clearly proper understanding causes underlying sensed world is key task-specific relevant information high-dimensional sets.

The construction of a meaningful graph plays crucial role in the success many graph-based representations and algorithms for handling structured data, especially emerging field signal processing. However, is not always readily available from nor easy to define depending on application domain. In particular, it often desirable processing applications that chosen such data admit certain regularity or smoothness graph. this paper, we address problem learning Laplacians, which equivalent...

Data augmentation is the process of generating samples by transforming training data, with target improving accuracy and robustness classifiers. In this paper, we propose a new automatic adaptive algorithm for choosing transformations used in data augmentation. Specifically, each sample, our main idea to seek small transformation that yields maximal classification loss on transformed sample. We employ trust-region optimization strategy, which consists solving sequence linear programs. Our...

State-of-the-art classifiers have been shown to be largely vulnerable adversarial perturbations. One of the most effective strategies improve robustness is training. In this paper, we investigate effect training on geometry classification landscape and decision boundaries. We show in particular that leads a significant decrease curvature loss surface with respect inputs, leading drastically more "linear" behaviour network. Using locally quadratic approximation, provide theoretical evidence...

Deep neural networks have recently shown impressive classification performance on a diverse set of visual tasks. When deployed in real-world (noise-prone) environments, it is equally important that these classifiers satisfy robustness guarantees: small perturbations applied to the samples should not yield significant loss predictor. The goal this article discuss deep may affect practice, including adversarial perturbations, random noise, and geometric transformations. This further discusses...

The construction of a meaningful graph topology plays crucial role in the effective representation, processing, analysis and visualization structured data. When natural choice is not readily available from data sets, it thus desirable to infer or learn In this tutorial overview, we survey solutions problem learning, including classical viewpoints statistics physics, more recent approaches that adopt signal processing (GSP) perspective. We further emphasize conceptual similarities differences...

Relationships between entities in datasets are often of multiple nature, like geographical distance, social relationships, or common interests among people a network, for example. This information can naturally be modeled by set weighted and undirected graphs that form global multilayer graph, where the vertex represents edges on different layers capture similarities term modalities. In this paper, we address problem analyzing multi-layer propose methods clustering vertices efficiently...

Deep Neural Networks have achieved extraordinary results on image classification tasks, but been shown to be vulnerable attacks with carefully crafted perturbations of the input data. Although most usually change values many image's pixels, it has that deep networks are also sparse alterations input. However, no computationally efficient method proposed compute perturbations. In this paper, we exploit low mean curvature decision boundary, and propose SparseFool, a geometry inspired attack...

Abstract Fragment-based drug discovery has been an effective paradigm in early-stage development. An open challenge this area is designing linkers between disconnected molecular fragments of interest to obtain chemically relevant candidate molecules. In work, we propose DiffLinker, E(3)-equivariant three-dimensional conditional diffusion model for linker design. Given a set fragments, our places missing atoms and designs molecule incorporating all the initial fragments. Unlike previous...

This paper addresses the problem of compression 3D point cloud sequences that are characterized by moving positions and color attributes. As temporally successive frames similar, motion estimation is key to effective these sequences. It however remains a challenging as have varying numbers points without explicit correspondence information. We represent time-varying geometry with set graphs, consider attributes clouds signals on vertices graphs. then cast feature matching between The...

Information analysis of data often boils down to properly identifying their hidden structure. In many cases, the structure can be described by a graph representation that supports signals in dataset. some applications, this may partly determined design constraints or predetermined sensing arrangements. general though, is not readily available nor easily defined. paper, we propose represent structured as sparse combination localized functions live on graph. This model more appropriate local...

Observational data usually comes with a multimodal nature, which means that it can be naturally represented by multi-layer graph whose layers share the same set of vertices (objects) different edges (pairwise relationships). In this paper, we address problem combining for an improved clustering compared to using independently. We propose two novel methods, are based on joint matrix factorization and regularization framework respectively, efficiently combine spectrum multiple layers, namely...

Unions of graph Fourier multipliers are an important class linear operators for processing signals defined on graphs. We present a novel method to efficiently distribute the application these high-dimensional collected by sensor networks. The proposed features approximations shifted Chebyshev polynomials, whose recurrence relations make them readily amenable distributed computation. demonstrate how can be used in denoising task, and show that communication requirements scale gracefully with...

Caching at mobile edge servers can smooth temporal traffic variability and reduce the service load of base stations in video delivery. However, assignment multiple representations to distributed is still a challenging question context adaptive streaming, since any two from different videos or even same will compete for limited caching storage. Therefore, it important, yet challenging, optimally select cached each server order effectively station while maintaining high quality experience...

In recent years Deep Neural Networks (DNNs) have been rapidly developed in various applications, together with increasingly complex architectures. The performance gain of these DNNs generally comes high computational costs and large memory consumption, which may not be affordable for mobile platforms. model quantization can used reducing the computation DNNs, deploying on equipment. this work, we propose an optimization framework deep quantization. First, a measurement to estimate effect...

Deep convolutional neural networks have been shown to be vulnerable arbitrary geometric transformations. However, there is no systematic method measure the invariance properties of deep such We propose ManiFool as a simple yet scalable algorithm networks. In particular, our measures robustness transformations in worst-case regime they can problematic for sensitive applications. Our extensive experimental results show that used fairly complex on high dimensional datasets and these values...

The goal of this paper is to analyze the geometric properties deep neural network image classifiers in input space. We specifically study topology classification regions created by networks, as well their associated decision boundary. Through a systematic empirical study, we show that state-of-the-art nets learn connected regions, and boundary vicinity datapoints flat along most directions. further draw an essential connection between two seemingly unrelated networks: sensitivity additive...