Publications

2022

• Assessing Vulnerability from Its Description
  
  • Zhang Zijing
  • Kumar Vimal
  • Mayo Michael
  • Bifet Albert
  , 2022, 1768, pp.129--143. This paper shows an end-to-end Artificial Intelligence (AI) system to estimate the severity level and the various Common Vulnerability Scoring System (CVSS) components from natural language descriptions without reproducing the vulnerability. This natural language processing-based approach can estimate the CVSS from only the Common Vulnerabilities and Exposures description without the need to reproduce the vulnerability environment. We present an Error Grid Analysis for the CVSS base score prediction task. Experiments on CVSS 2.0 and CVSS 3.1 show that state-of-the-art deep learning models can predict the CVSS scoring components with high accuracy. The low-cost Universal Sentence Encoder (large) model outperforms the Generative Pre-trained Transformer-3 (GPT-3) and the Support Vector Machine baseline on the majority of the classification tasks with a lower computation overhead than the GPT-3. (10.1007/978-981-99-0272-9_9)
  DOI : 10.1007/978-981-99-0272-9_9
• Interleaved Challenge Loop PUF: A Highly Side-Channel Protected Oscillator-Based PUF
  
  • Tebelmann Lars
  • Danger Jean-Luc
  • Pehl Michael
  IEEE Transactions on Circuits and Systems I: Regular Papers, IEEE, 2022, 69 (12), pp.5121-5134. Physical Unclonable Functions (PUFs) leverage manufacturing variations to generate device-specific keys during runtime only, overcoming the need for protection after power-off as for Non-Volatile Memory. The main challenges of PUF-based key storage are reliability of the response and sensitivity to Side-Channel Analysis (SCA). Oscillator-based PUFs are particularly sensitive to frequency spectrum SCA. Existing countermeasures can protect sign-based bit derivation that requires error correction or discarding unreliable bits to achieve reliable key generation. Amplitude-based bit derivation enhances the reliability of oscillator-based PUFs without discarding unsteady response bits, keeping a high entropy. However, existing lightweight countermeasures are not applicable for this case. This raises the demand for an alternative solution. This work targets the protection of amplitude-based bit derivation combined with the Loop PUF, an oscillator-based PUF primitive well suited for key generation. It presents the Interleaved Challenge Loop PUF (ICLooPUF), a side-channel-hardened offspring of the Loop PUF that uses dynamic challenge interleaving. The SCA-protected PUF primitive is applicable to amplitude-based and sign-based bit derivation methods, and requires a low hardware overhead. Theoretical and experimental results show the efficiency and effectiveness of the protection mechanism. (10.1109/TCSI.2022.3208325)
  DOI : 10.1109/TCSI.2022.3208325
• Membership Inference Attacks via Adversarial Examples
  
  • Jalalzai Hamid
  • Kadoche Elie
  • Leluc Rémi
  • Plassier Vincent
  , 2022. The raise of machine learning and deep learning led to significant improvement in several domains. This change is supported by both the dramatic rise in computation power and the collection of large datasets. Such massive datasets often include personal data which can represent a threat to privacy. Membership inference attacks are a novel direction of research which aims at recovering training data used by a learning algorithm. In this paper, we develop a mean to measure the leakage of training data leveraging a quantity appearing as a proxy of the total variation of a trained model near its training samples. We extend our work by providing a novel defense mechanism. Our contributions are supported by empirical evidence through convincing numerical experiments.
• A 40 MHz 11-Bit ENOB Delta Sigma ADC for Communication and Acquisition Systems
  
  • Fakhoury Hussein
  • Jabbour Chadi
  • Nguyen Van-Tam
  Sensors, MDPI, 2022, 23 (1), pp.36. This paper describes a Delta Sigma ADC IC that embeds a 5th-order Continuous-Time Delta Sigma modulator with 40 MHz signal bandwidth, a low ripple 20 to 80 MS/s variable-rate digital decimation filter, a bandgap voltage reference, and high-speed CML buffers on a single die. The ADC also integrates on-chip calibrations for RC time-constant variation and quantizer offset. The chip was fabricated in a 1P7M 65 nm CMOS process. Clocked at 640 MHz, the Continuous-Time Delta Sigma modulator achieves 11-bit ENOB and 76.5 dBc THD up to 40 MHz of signal bandwidth while consuming 82.3 mW. (10.3390/s23010036)
  DOI : 10.3390/s23010036
• A dynamic attack graphs based approach for impact assessment of vulnerabilities in complex computer systems
  
  • Boudermine Antoine
  , 2022. Nowadays, computer networks are used in many fields and their breakdown can strongly impact our daily life. Assessing their security is a necessity to reduce the risk of compromise by an attacker. Nevertheless, the solutions proposed so far are rarely adapted to the high complexity of modern computer systems. They often rely on too much human work and the algorithms used don't scale well. Furthermore, the evolution of the system over time is rarely modeled and is therefore not considered in the evaluation of its security.In this thesis, we propose a new attack graph model built from a dynamic description of the system. We have shown through our experimentations that our model allows to identify more attack paths than a static attack graph model. We then proposed an attack simulation algorithm to approximate the chances of success of system compromise by a malicious actor.We also proved that our solution was able to analyze the security of complex systems. The worst-case time complexity was assessed for each algorithm used. Several tests were performed to measure their real performances. Finally, we applied our solution on an IT network composed of several thousand elements.Future work should be done to improve the performance of the attack graph generation algorithm in order to analyze increasingly complex systems. Solutions should also be found to facilitate the system modeling step which is still a difficult task to perform, especially by humans. Finally, the simulation algorithm could be improved to be more realistic and take into account the real capabilities of the attacker. It would also be interesting to assess the impact of the attacks on the organization and its business processes.
• Low-Latency Sliding Window Algorithms for Formal Languages
  
  • Ganardi Moses
  • Jachiet Louis
  • Lohrey Markus
  • Schwentick Thomas
  , 2022. Low-latency sliding window algorithms for regular and context-free languages are studied, where latency refers to the worst-case time spent for a single window update or query. For every regular language $L$ it is shown that there exists a constant-latency solution that supports adding and removing symbols independently on both ends of the window (the so-called two-way variable-size model). We prove that this result extends to all visibly pushdown languages. For deterministic 1-counter languages we present a $\mathcal{O}(\log n)$ latency sliding window algorithm for the two-way variable-size model where $n$ refers to the window size. We complement these results with a conditional lower bound: there exists a fixed real-time deterministic context-free language $L$ such that, assuming the OMV (online matrix vector multiplication) conjecture, there is no sliding window algorithm for $L$ with latency $n^{1/2-\epsilon}$ for any $\epsilon>0$, even in the most restricted sliding window model (one-way fixed-size model). The above mentioned results all refer to the unit-cost RAM model with logarithmic word size. For regular languages we also present a refined picture using word sizes $\mathcal{O}(1)$, $\mathcal{O}(\log\log n)$, and $\mathcal{O}(\log n)$. (10.4230/LIPIcs.FSTTCS.2022.38)
  DOI : 10.4230/LIPIcs.FSTTCS.2022.38
• Continuous Health Monitoring of Machinery using Online Clustering on Unlabeled Data Streams
  
  • Nguyen Minh-Huong Le
  • Turgis Fabien
  • Fayemi Pierre-Emmanuel
  • Bifet Albert
  , 2022, pp.1866--1873. Maintenance is an important support function to ensure the reliability, safety, and availability in the railway. Lately, machine learning has become a major player and allows practitioners to build intricate learning models for machinery maintenance. Commonly, a model is trained on static data and is retrained on new data that exhibit novelties unknown to the model. On the contrary, online machine learning is a learning paradigm that adapts the models to new data, thus enabling adaptive, lifelong learning. Our goal is to leverage online learning on unlabeled data streams to enhance railway machinery maintenance. We propose Continuous Health Monitoring using Online Clustering (CheMoc) as an unsupervised method that learns the health profiles of the systems incrementally, assesses their working condition continuously via an adaptive health score, and works efficiently on streaming data. We evaluate CheMoc on a real-world data set from a national railway company. The results show that CheMoc discovered relevant health clusters, as confirmed by a domain expert, and processed the data of an entire year under two hours using only 600 MB of memory. (10.1109/BIGDATA55660.2022.10021002)
  DOI : 10.1109/BIGDATA55660.2022.10021002
• Stream2Graph: Dynamic Knowledge Graph for Online Learning Applied in Large-scale Network
  
  • Barry Mariam
  • Bifet Albert
  • Chiky Raja
  • Jaouhari Saad El
  • Montiel Jacob
  • Ouafi Aissa El
  • Guerizec Eric
  , 2022, pp.2190--2197. Knowledge Graphs (KG) are valuable information sources that store knowledge in a domain (healthcare, finance, e-commerce, cyber-security.). Most industrial KGs are dynamic by nature as they are updated regularly with streaming data (customer activity, network traffic, application logs, IT process). However, extracting insights from continuously updated data comes with major challenges, particularly in big data settings. In this paper, we address the following challenges: 1) ingesting heterogeneous data, 2) training and deployment of predictive models on continuously evolving data, and 3) implementation of data pipelines for updating and maintaining the KG in production. We cover multiple aspects of this process, from knowledge collection to its operationalization. We propose Stream2Graph, a stream-based system for building and updating the knowledge base dynamically in real time. Then we show how graph features can be used in downstream online machine learning models. The solution speeds up big data stream learning and knowledge extraction to enhance Graph-based AI applications. Experimental results show the effectiveness of our solution for knowledge base construction and improvement of big data learning capabilities. Using data from Stream2Graph resulted in speedups for training and inference time in the range from 547x to 2000x in downstream ML models. Finally, we provide the lessons learned from applying graph-based online learning on large-scale network processing high-velocity streaming data. (10.1109/BIGDATA55660.2022.10020885)
  DOI : 10.1109/BIGDATA55660.2022.10020885
• StreamFlow: A System for Summarizing and Learning Over Industrial Big Data Streams
  
  • Barry Mariam
  • Jaouhari Saad El
  • Bifet Albert
  • Montiel Jacob
  • Guerizec Eric
  • Chiky Raja
  , 2022, pp.2198--2205. The growing need for predictive analytics over streaming data in the industry requires a flexible and continuously scalable big data system. In real-time big data applications (cybersecurity, AIOps, anomaly detection, predictive maintenance, IoT etc.), efficient machine learning models must be trained and industrialized within existing data processing plat-forms and industrial tools. This requires interoperability between various components: data collection, processing, summarization, modelling and analytics. Existing works focus on building AI models for big data, neglecting real-world challenges when integrating such models into an existing industrial production framework. In this paper, we propose StreamFlow, an operational data pipeline to address industrial challenges for continuous learning over big data streams. We also propose an online method using sliding windows to summarize high-velocity data. The final result of the framework is a feature vector that describes the underlying processes and is ready to use in machine learning tasks. Moreover, we showcase real-world applications such as automated feature engineering for real-time monitoring and online machine learning for event classification. The proposed system has been deployed within production in a banking system, processing billions of daily traffic operations. Our experiments demonstrate the effectiveness and performance of our approach by evaluating it at different levels: processing, summarization, improvement of machine learning performance and effectiveness in an industrial setting. In the case of downstream machine learning tasks, using summarized data generated by StreamFlow results in up to 2 orders of magnitude speedups in training time without compromising predictive performance. (10.1109/BIGDATA55660.2022.10020438)
  DOI : 10.1109/BIGDATA55660.2022.10020438
• Morphing and level-of-detail operators for interactive digital material design and rendering
  
  • Gauthier Alban
  , 2022. The Physically Based Rendering workflow has become a standard for rendering digital materials for the creative industries, such as video games, visual special effects, product design and architecture. It enables developers and artists to create and share ready-to-use photorealistic materials among a wide variety of applications.In this workflow, 3D surfaces are mapped to a 2D texture space where their Spatially Varying Bidirectional Reflectance Distribution Functions are encoded as a set of bitmap images called PBR maps queried efficiently at runtime. These maps represent interpretable physically based quantities while allowing for the reproduction of a wide range of material appearances. They can be reconstructed from real-world photographs or generated procedurally.Unfortunately, both approaches to PBR material authoring require advanced skills and a significant amount of time to model convincing materials to be used by photorealistic renderers. In addition, while all channels are encoded in the same pixel grid, they describe heterogeneous quantities of very different nature at different scales that are partly correlated. The information described in the maps can be either geometrical for the height, normal, and roughness or colorimetric for the albedo. The roughness relates to the distribution of microfacet normals, embedded atop the normal's tangent plane, which location is given by the height map. This description allows for efficient renderings but prevents the use of simple image processing operators jointly across maps for interpolating or filtering.In this thesis, we explore efficient morphing and level-of-detail operators to tackle these difficulties. We propose a novel morphing operator which allows creating new materials by simply blending two existing ones while preserving their dominant structures and features all along the interpolation. This operator allows exploring large regions of the space of possible materials using exemplars as anchors and our interpolation scheme as a navigation means. We also propose a novel approach for SVBRDF mipmapping which preserves material appearance under varying view distances and lighting conditions. As a result, we obtain a drop-in replacement for standard material mipmapping, offering a significant improvement in appearance preservation while still boiling down to a single per-pixel mipmap texture fetch. These operators have been experimentally validated on a large dataset of examples.Overall, our proposed methods allow for interpolating materials in the canonical space of textures as well as along the downscaling pyramid for preserving and exploring appearance.
• Socio-conversational systems: Three challenges at the crossroads of fields
  
  • Clavel Chloé
  • Labeau Matthieu
  • Cassell Justine
  Frontiers in Robotics and AI, Frontiers Media S.A., 2022, 9. Socio-conversational systems are dialogue systems, including what are sometimes referred to as chatbots, vocal assistants, social robots, and embodied conversational agents, that are capable of interacting with humans in a way that treats both the specifically social nature of the interaction and the content of a task. The aim of this paper is twofold: 1) to uncover some places where the compartmentalized nature of research conducted around socio-conversational systems creates problems for the field as a whole, and 2) to propose a way to overcome this compartmentalization and thus strengthen the capabilities of socio-conversational systems by defining common challenges. Specifically, we examine research carried out by the signal processing, natural language processing and dialogue, machine/deep learning, social/affective computing and social sciences communities. We focus on three major challenges for the development of effective socio-conversational systems, and describe ways to tackle them. (10.3389/frobt.2022.937825)
  DOI : 10.3389/frobt.2022.937825
• Interpretable Generative Modeling Using a Hierarchical Topological VAE
  
  • Desticourt Etienne
  • Letort Véronique
  • d'Alché-Buc Florence
  , 2022, pp.1415-1421. Generating realistic datasets with fine-grained control over their properties can help overcome challenges linked to the scarcity of data in many domains, such as medical applications. To that end, we extend Variational Autoencoders by using a hierarchical and topological prior consisting of a sequence of Self-Organizing Maps (SOM), which are stacked in the latent space and learned without supervision, jointly with the parameters of the variational autoencoder. We induce a hierarchy between the codes of the SOM sequence, each SOM corresponding to a different hierarchical level and learning increasingly finer-grained representations of the data. Our model combines the power of deep learning with the interpretability of hierarchical and topological clustering and produces competitive results when evaluated on three well-known computer vision benchmarks and a custom medical dataset. (10.1109/CSCI58124.2022.00253)
  DOI : 10.1109/CSCI58124.2022.00253
• [Tutorial] Linear Video Coding and Transmission Schemes for Next Generation Video Applications
  
  • Trioux Anthony
  • Coudoux François-Xavier
  • Cagnazzo Marco
  • Kieffer Michel
  , 2022.
• Broadcast Encryption and Traitor Tracing
  
  • Phan Duong Hieu
  , 2022 (1). Broadcast encryption and revocation schemes can be covered by more general primitives such as attribute-based encryption. However, when applying a general framework to a concrete primitive, this often results in impractical schemes. This chapter talks about recent advancements of this approach. A multi-receiver encryption scheme with the ability to trace traitors is called a traitor tracing (TT). The chapter presents an overview of the different techniques for designing broadcast encryption and TT. Combinatorial broadcast encryption schemes are mainly based on a tree structure or on a fingerprinting code. Tree-based schemes support revocation but have limited capacity dealing with tracing traitors, while code-based ones provide traceability but very few support revocation. Fingerprinting with collusion-secure codes allows one to identify a digital document among several copies of it by embedding a fingerprint. (10.1002/9781394188369.ch6)
  DOI : 10.1002/9781394188369.ch6
• Shelter Check: Proactively Finding Tax Minimization Strategies via AI
  
  • Blair-Stanek Andrew
  • Holzenberger Nils
  • van Durme Benjamin
  Tax Notes Federal, Tax Analysts, 2022, 177. In this article, the authors explore how artificial intelligence could be used to automatically find tax minimization strategies in the tax law. Congress or Treasury could then proactively shut down such strategies. But, if large accounting or law firms develop the technology first, the result could be a huge, silent hit to the treasury.
• Reinforcement Learning Based Architectures for Dynamic Generation of Smart Home Services
  
  • Qiu Mingming
  • Najm Elie
  • Sharrock Rémi
  • Traverson Bruno
  , 2022, pp.7-14. A smart home system is realized by implementing various services. However, the design and deployment of smart home services are challenging due to their complexity and the large number of connected objects. Existing approaches to the smart home system to create services either require complex input from the inhabitant or can only work if the inhabitant specifies regulation solutions rather than targets. In addition, smart home services may conflict if they access the same actuators. Learning methods to dynamically generate smart home services are promising ways to solve the above problems. In this paper, depending on the ability to consider the composition of services and their mutual influence, we propose several reinforcement learning-based architectures for a smart home system to dynamically generate services. The expected advantages are, first, that the smart home services can propose the states of the actuators by considering the target values of the controllable environment states given by the inhabitant or by interacting with the inhabitant in a simple and natural way; and second, that there is no conflict between these propositions. We compare the performance of the proposed architectures using several simulated smart home environments with different services and select the architectures with the best performance concerning our predefined metrics. (10.1109/ICMLA55696.2022.00010)
  DOI : 10.1109/ICMLA55696.2022.00010
• Practical homomorphic evaluation of block-cipher-based hash functions with applications
  
  • Bendoukha Adda-Akram
  • Stan Oana
  • Sirdey Renaud
  • Quero Nicolas
  • Freitas Luciano
  , 2022, 13877, pp.88-103. Fully homomorphic encryption (FHE) is a powerful cryptographic technique allowing to perform computation directly over encrypted data. Motivated by the overhead induced by the homomorphic ciphertexts during encryption and transmission, the transciphering technique, consisting in switching from a symmetric encryption to FHE encrypted data was investigated in several papers. Different stream and block ciphers were evaluated in terms of their "FHE-friendliness", meaning practical implementations costs while maintaining sufficient security levels. In this work, we present a first evaluation of hash functions in the homomorphic domain, based on well-chosen block ciphers. More precisely, we investigate the cost of transforming PRINCE and SIMON, two lightweight block-ciphers into secure hash functions using well-established block-cipher-based hash functions constructions, and provide evaluation under bootstrappable FHE schemes. We also motivate the necessity of practical homomorphic evaluation of hash functions by providing several use cases in which the integrity of private data is also required. In particular, our hash constructions can be of significant use in a threshold-homomorphic based protocol for the single secret leader election problem occuring in blockchains with Proof-of-stake consensus. Our experiments showed that using a TFHE implementation of a hash function, we are able to achieve practical runtime, and appropriate security levels. (10.1007/978-3-031-30122-3_6)
  DOI : 10.1007/978-3-031-30122-3_6
• Entropic Hardness of Module-LWE from Module-NTRU
  
  • Boudgoust Katharina
  • Jeudy Corentin
  • Roux-Langlois Adeline
  • Wen Weiqiang
  , 2022, 13774, pp.78 - 99. The Module Learning With Errors problem (M-LWE) has gained popularity in recent years for its security-efficiency balance, and its hardness has been established for a number of variants. In this paper, we focus on proving the hardness of (search) M-LWE for general secret distributions, provided they carry sufficient min-entropy. This is called entropic hardness of M-LWE. First, we adapt the line of proof of Brakerski and Döttling on R-LWE (TCC'20) to prove that the existence of certain distributions implies the entropic hardness of M-LWE. Then, we provide one such distribution whose required properties rely on the hardness of the decisional Module-NTRU problem. (10.1007/978-3-031-22912-1_4)
  DOI : 10.1007/978-3-031-22912-1_4
• Stochastic Model of Sub-Poissonian Quantum Light in an Interband Cascade Laser
  
  • Zhao Shiyuan
  • Grillot Frédéric
  Physical Review Applied, American Physical Society, 2022, 18. This work theoretically investigates the possibility of generating amplitude-squeezed light with high-quantum-efficiency interband cascade lasers. Based on a stochastic approach, we show that, by employing the suppressed-pump-noise configuration, this kind of midinfrared source enables operation with considerable amplitude squeezing over a large bandwidth of several GHz. Our results facilitate future midinfrared quantum photonic applications such as free-space secure communications. (10.1103/physrevapplied.18.064027)
  DOI : 10.1103/physrevapplied.18.064027
• An approach to bridge ROS 1 and ROS 2 devices into an OPC UA-based testbed for industry 4.0
  
  • Nguyen Quang-Duy
  • Dhouib Saadia
  • Huang Yining
  • Bellot Patrick
  , 2022. ROS 1 and ROS 2 are two widely-used robotic middleware. One of their essential features is to enable two robots with the same middleware, ROS 1 or ROS 2, to directly connect and collaborate. However, two robots running two different middleware can only communicate by additionally using one of the bridge solutions available in the robotic community. It is even more challenging when deploying these robots as part of an OPC UA-based industrial testbed. The first challenge is to network the robots with other OPC UA devices. Second, a testbed environment sometimes requires a robot to join the system rapidly and with minimal configuration for quick experiments. While addressing the above needs, this paper presents an approach to bridge ROS 1 and ROS 2 robots to an OPC UA PubSub network. The approach derives from the actual experiences in developing an OPC UAbased robotic testbed for Industry 4.0 research.
• Assessing Performance and Fairness Metrics in Face Recognition – Bootstrap Methods
  
  • Conti Jean-Rémy
  • Clémençon Stéphan
  , 2022. The ROC curve is the major tool for assessing not only the performance but also the fairness properties of a similarity scoring function in Face Recognition. In order to draw reliable conclusions based on empirical ROC analysis, evaluating accurately the uncertainty related to statistical versions of the ROC curves of interest is necessary. For this purpose, we explain in this paper that, because the True/False Acceptance Rates are of the form of U-statistics in the case of similarity scoring, the naive bootstrap approach is not valid here and that a dedicated recentering technique must be used instead. This is illustrated on real data of face images, when applied to several ROC-based metrics such as popular fairness metrics.
• An Adversarial Robustness Perspective on the Topology of Neural Networks
  
  • Goibert Morgane
  • Ricatte Thomas
  • Dohmatob Elvis
  , 2022. In this paper, we investigate the impact of neural networks (NNs) topology on adversarial robustness. Specifically, we study the graph produced when an input traverses all the layers of a NN, and show that such graphs are different for clean and adversarial inputs. We find that graphs from clean inputs are more centralized around highway edges, whereas those from adversaries are more diffuse, leveraging under-optimized edges. Through experiments on a variety of datasets and architectures, we show that these under-optimized edges are a source of adversarial vulnerability and that they can be used to detect adversarial inputs.
• Approach for Early Validation of System Requirements Application to software architecture of autonomous vehicles
  
  • Assioua Yasmine
  • Ameur-Boulifa Rabea
  • Guitton-Ouhamou Patricia
  • Pacalet Renaud
  , 2022. The automotive industry is changing, digital components are gradually augmenting or replacing mechanical systems. The advent of autonomous and connected cars further increases the number and the complexity of embedded electronic systems, which poses new challenges. Indeed, compared to conventional vehicles, these highly technological objects have an increased role in the safety of their passengers and their environment. The requirements in terms of reliability and safety are thus also increased. To approach this new era, manufacturers must adapt their methods and invent new ones. The thesis proposes a method to meet some reliability and security related challenges that the limitations of the traditional approaches do not solve properly. It consists in introducing validation as early as possible in the software design and development life cycle, even before any executable or formal model has been produced. The method lays foundations for an iterative approach for the validation and verification of textual requirements and statements in order to detect errors, omissions or inconsistencies before implementation. This requirement qualification process is based on the translation of informal descriptions into formal models, followed by formal verification of generic or custom temporal logic properties. It also uses simulations and traces analysis for counter-example investigation. The proposed method is largely automated.
• TINA: Textual Inference with Negation Augmentation
  
  • Helwe Chadi
  • Coumes Simon
  • Clavel Chloé
  • Suchanek Fabian
  , 2022. Transformer-based language models achieve state-of-the-art results on several natural language processing tasks. One of these is textual entailment, i.e., the task of determining whether a premise logically entails a hypothesis. However, the models perform poorly on this task when the examples contain negations. In this paper, we propose a new definition of textual entailment that captures also negation. This allows us to develop TINA (Textual Inference with Negation Augmentation), a principled technique for negated data augmentation that can be combined with the unlikelihood loss function. Our experiments with different transformer-based models show that our method can significantly improve the performance of the models on textual entailment datasets with negation-without sacrificing performance on datasets without negation.
• Depth Patch Selection for Decoder-Side Depth Estimation in MPEG Immersive Video
  
  • Milovanovic Marta
  • Henry Felix
  • Cagnazzo Marco
  , 2022, pp.343-347. The MPEG immersive video (MIV) standard has been developed to efficiently compress volumetric video content and enable an immersive user experience. MIV deals with an enormous amount of data that comes in the form of multi-view plus depth videos, which is efficiently reduced in the process of pruning, by tackling the redundancies among the views. This paper presents a novel approach for improving the existing immersive video coding scheme. The proposed approach reduces the amount of transmitted depth data, leveraging the fact that the depth information is partially contained in texture videos. The study proposes a method that ensures a reliable recovery of depths at the decoder-side. This method provides BD-rate improvements on both high and low bitrate ranges, with up to 22.57% Y-PSNR, 25.76%VMAF, 24.07% MS-SSIM, and 22.94% IV-PSNR metric gain, given a low bitrate setting. (10.1109/PCS56426.2022.10018042)
  DOI : 10.1109/PCS56426.2022.10018042