Publications

2023

• Efficient learning of the structure and parameters of local Pauli noise channels
  
  • Rouzé Cambyse
  • Stilck Franca Daniel
  , 2023. The unavoidable presence of noise is a crucial roadblock for the development of large-scale quantum computers and the ability to characterize quantum noise reliably and efficiently with high precision is essential to scale quantum technologies further. Although estimating an arbitrary quantum channel requires exponential resources, it is expected that physically relevant noise has some underlying local structure, for instance that errors across different qubits have a conditional independence structure. Previous works showed how it is possible to estimate Pauli noise channels with an efficient number of samples in a way that is robust to state preparation and measurement errors, albeit departing from a known conditional independence structure. We present a novel approach for learning Pauli noise channels over n qubits that addresses this shortcoming. Unlike previous works that focused on learning coefficients with a known conditional independence structure, our method learns both the coefficients and the underlying structure. We achieve our results by leveraging a groundbreaking result by Bresler for efficiently learning Gibbs measures and obtain an optimal sample complexity of O(log(n)) to learn the unknown structure of the noise acting on n qubits. This information can then be leveraged to obtain a description of the channel that is close in diamond distance from O(poly(n)) samples. Furthermore, our method is efficient both in the number of samples and postprocessing without giving up on other desirable features such as SPAM-robustness, and only requires the implementation of single qubit Cliffords. In light of this, our novel approach enables the large-scale characterization of Pauli noise in quantum devices under minimal experimental requirements and assumptions. (10.48550/arXiv.2307.02959)
  DOI : 10.48550/arXiv.2307.02959
• (Adversarial) Electromagnetic Disturbance in the Industry
  
  • Beckers Arthur
  • Guilley Sylvain
  • Maurine Philippe
  • O'Flynn Colin
  • Picek Stjepan
  IEEE Transactions on Computers, Institute of Electrical and Electronics Engineers, 2023, 72 (2), pp.414-422. Faults occur naturally and are responsible for reliability concerns. Faults are also an interesting tool for attackers to extract sensitive information from secure chips. In particular, non-invasive fault attacks have received a fair amount of attention. One easy way to perturb a chip without altering it is the so-called Electromagnetic Fault Injection (EMFI). Such attack has been studied in great depth, and nowadays, it is part and parcel of the state-of-the-art. Indeed, new capabilities have emerged where EM experimental benches are used to cryptanalyze chips. The progress of this "field" is fast, in terms of reproducibility, accuracy, and number of use-cases. However, there is too little awareness about such advances. In this paper, we aim to expose the true harmfulness of EMFI (including reproducibility) to enable reasonable security quotations. We also analyze protections (at hardware/firmware/system levels) in light of their efficiency. We characterize the specificity of EM fault injection compared to other injection means (laser, glitch, probing). (10.1109/TC.2022.3224373)
  DOI : 10.1109/TC.2022.3224373
• Locally differentially private estimation of nonlinear functionals of discrete distributions
  
  • Butucea Cristina
  • Issartel Yann
  Advances in Neural Information Processing Systems, Morgan Kaufmann Publishers, 2023. We study the problem of estimating non-linear functionals of discrete distributions in the context of local differential privacy. The initial data $x_1,\ldots,x_n \in [K]$ are supposed i.i.d. and distributed according to an unknown discrete distribution $p = (p_1,\ldots,p_K)$. Only $\alpha$-locally differentially private (LDP) samples $z_1,...,z_n$ are publicly available, where the term 'local' means that each $z_i$ is produced using one individual attribute $x_i$. We exhibit privacy mechanisms (PM) that are interactive (i.e. they are allowed to use already published confidential data) or non-interactive. We describe the behavior of the quadratic risk for estimating the power sum functional $F_{\gamma} = \sum_{k=1}^K p_k^{\gamma}$, $\gamma >0$ as a function of $K, \, n$ and $\alpha$. In the non-interactive case, we study two plug-in type estimators of $F_{\gamma}$, for all $\gamma >0$, that are similar to the MLE analyzed by Jiao et al. (2017) in the multinomial model. However, due to the privacy constraint the rates we attain are slower and similar to those obtained in the Gaussian model by Collier et al. (2020). In the interactive case, we introduce for all $\gamma >1$ a two-step procedure which attains the faster parametric rate $(n \alpha^2)^{-1/2}$ when $\gamma \geq 2$. We give lower bounds results over all $\alpha$-LDP mechanisms and all estimators using the private samples.
• Scaling by subsampling for big data, with applications to statistical learning
  
  • Bertail Patrice
  • Bouchouia Mohammed
  • Jelassi Ons
  • Tressou Jessica
  • Zetlaoui Mélanie
  Journal of Nonparametric Statistics, American Statistical Association, 2023, 36 (1), pp.78-117. Handling large datasets and calculating complex statistics on huge datasets require important computing resources. Using subsampling methods to calculate statistics of interest on small samples is often used in practice to reduce computational complexity, for instance using the divide and conquer strategy. In this article, we recall some results on subsampling distributions and derive a precise rate of convergence for these quantities and the corresponding quantiles. We also develop some standardization techniques based on subsampling unstandardized statistics in the framework of large datasets. It is argued that using several subsampling distributions with different subsampling sizes brings a lot of information on the behavior of statistical learning procedures: subsampling allows to estimate the rate of convergence of different algorithms, to estimate the variability of complex statistics, to estimate confidence intervals for out-of-sample errors and interpolate their values at larger scales. These results are illustrated on simulations, but also on two important datasets, frequently analyzed in the statistical learning community, EMNIST (recognition of digits) and VeReMi (analysis of Network Vehicular Reference Misbehavior). (10.1080/10485252.2023.2219782)
  DOI : 10.1080/10485252.2023.2219782
• Multi-Agent Systems
  
  • Malvone Vadim
  • Murano Aniello
  , 2023, 14282, pp.XX-554. This volume LNCS 14282 constitutes the refereed proceedings of the 20th European Conference EUMAS 2023, held in Naples, Italy, during September 2023. This volume includes 24 full papers and 5 short papers, carefully selected from 47 submissions. Additionally, the volume features 16 short papers, rigorously reviewed from 20 submissions for the PhD day. The conference focused on the theory and practice of autonomous agents and multi-agent systems, covering a wide range of topics. (10.1007/978-3-031-43264-4)
  DOI : 10.1007/978-3-031-43264-4
• Execution trace analysis for a precise understanding of latency violations
  
  • Zoor Maysam
  • Apvrille Ludovic
  • Pacalet Renaud
  • Coudert Sophie
  Software and Systems Modeling, Springer Verlag, 2023. Despite the amount of proposed works for the verification of embedded systems, understanding the root cause of violations of requirements in simulation or execution traces is still an openissue, especially when dealing with temporal properties such as latencies. Is the violation due to an unfavorable real-time scheduling, to contentions on buses, to the characteristics of functional algorithms or hardware components? The paper introduces the Precise Latency ANalysis approach (PLAN), a new trace analysis technique whose objective is to classify execution transactions according to their impact on latency. To do so, we rely first on a model transformation that builds up a dependency graph from an allocation model, thus including hardware and software aspects of a system model. Then, from this graph and an execution trace, our analysis can highlight how software or hardware elements contributed to the latency violation. The paper first formalizes the problem before applying our approach to simulation traces of SysML models. A case study defined in the AQUAS European project illustrates the relevance of our approach. Last, a performance evaluation gives computation times for several models and requirements. (10.1007/s10270-022-01076-z)
  DOI : 10.1007/s10270-022-01076-z
• Phase estimation at the point-ahead angle for AO pre-compensated ground to GEO satellite telecoms
  
  • Lognoné Perrine
  • Conan Jean-Marc
  • Rekaya Ghaya
  • Védrenne Nicolas
  Optics Express, Optical Society of America - OSA Publishing, 2023, 31 (3), pp.3441. We present a new method to estimate the off-axis adaptive optics pre-compensation phase of a ground to GEO satellite telecom link suffering from point-ahead anisoplanatism. The proposed phase estimator relies on the downlink phase and log-amplitude measurements that are available at the optical ground station. We introduce the analytical tools, extended from the literature, to build the estimator as well as a general modal formalism to express the reciprocal residual phase covariance matrix resulting from any estimation linear with measurements. We use this residual phase covariance matrix to generate independent coupled flux samples thanks to a pseudo-analytical approach and study the gain offered by the proposed estimator on the coupled flux statistics, in various atmospheric conditions. The estimator is shown to reduce the anisoplanatic residual phase variance by at least 35%, and 46% at best, with a greater impact on the lower modes, especially on the tip and tilt residual phase variances. The phase variance reduction brings a gain up to 15 dB on the cumulative density function at probability 10 −3 . This gain should allow to relax the power constraints on the link budget at the OGS and renews the interest in large aperture diameter (60 cm class telescopes) for GEO Feeder links by reducing the atmospheric turbulence impact on the uplink coupled signal. (10.1364/OE.476328)
  DOI : 10.1364/OE.476328
• Contrast, Stylize and Adapt: Unsupervised Contrastive Learning Framework for Domain Adaptive Semantic Segmentation
  
  • Li Tianyu
  • Roy Subhankar
  • Zhou Huayi
  • Lu Hongtao
  • Lathuilière Stéphane
  , 2023. To overcome the domain gap between synthetic and real-world datasets, unsupervised domain adaptation methods have been proposed for semantic segmentation. Majority of the previous approaches have attempted to reduce the gap either at the pixel or feature level, disregarding the fact that the two components interact positively. To address this, we present CONtrastive FEaTure and pIxel alignment (CONFETI) for bridging the domain gap at both the pixel and feature levels using a unique contrastive formulation. We introduce well-estimated prototypes by including category-wise cross-domain information to link the two alignments: the pixel-level alignment is achieved using the jointly trained style transfer module with the prototypical semantic consistency, while the feature-level alignment is enforced to cross-domain features with the \textbf{pixel-to-prototype contrast}. Our extensive experiments demonstrate that our method outperforms existing state-of-the-art methods using DeepLabV2. Our code is available at https://github.com/cxa9264/CONFETI
• PrivacyGAN: robust generative image privacy
  
  • Zameshina Mariia
  • Careil Marlene
  • Teytaud Olivier
  • Najman Laurent
  , 2023. Classical techniques for protecting facial image privacy typically fall into two categories: data-poisoning methods, exemplified by Fawkes, which introduce subtle perturbations to images, or anonymization methods that generate images resembling the original only in several characteristics, such as gender, ethnicity, or facial expression. In this study, we introduce a novel approach, PrivacyGAN, that uses the power of image generation techniques, such as VQGAN and StyleGAN, to safeguard privacy while maintaining image usability, particularly for social media applications. Drawing inspiration from Fawkes, our method entails shifting the original image within the embedding space towards a decoy image. We evaluate our approach using privacy metrics on traditional and novel facial image datasets. Additionally, we propose new criteria for evaluating the robustness of privacy-protection methods against unknown image recognition techniques, and we demonstrate that our approach is effective even in unknown embedding transfer scenarios. We also provide a human evaluation that further proves that the modified image preserves its utility as it remains recognisable as an image of the same person by friends and family.
• Knowledge Bases and Language Models: Complementing Forces
  
  • Suchanek Fabian M.
  • Luu Anh Tuan
  , 2023. Large language models (LLMs), as a particular instance of generative articial intelligence, have revolutionized natural language processing. In this invited paper, we argue that LLMs are complementary to structured data repositories such as databases or knowledge bases, which use symbolic knowledge representations. Hence, the two ways of knowledge representation will likely continue to co-exist, at least in the near future. We discuss ways that have been explored to make the two approaches work together, and point out opportunities and challenges for their symbiosis.
• Using the Uniqueness of Global Identifiers to Determine the Provenance of Python Software Source Code
  
  • Sun Yiming
  • German Daniel M.
  • Zacchiroli Stefano
  Empirical Software Engineering, Springer Verlag, 2023. We consider the problem of identifying the provenance of free/open source software (FOSS) and specifically the need of identifying where reused source code has been copied from. We propose a lightweight approach to solve the problem based on software identifiers—such as the names of variables, classes, and functions chosen by programmers. The proposed approach is able to efficiently narrow down to a small set of candidate origin products, to be further analyzed with more expensive techniques to make a final provenance determination. By analyzing the PyPI (Python Packaging Index) open source ecosystem we find that globally defined identifiers are very distinct. Across PyPI's 244 K packages we found 11.2 M different global identifiers (classes and method/function names—with only 0.6% of identifiers shared among the two types of entities); 76% of identifiers were used only in one package, and 93% in at most 3. Randomly selecting 3 non-frequent global identifiers from an input product is enough to narrow down its origins to a maximum of 3 products within 89% of the cases. We validate the proposed approach by mapping Debian source packages implemented in Python to the corresponding PyPI packages; this approach uses at most five trials, where each trial uses three randomly chosen global identifiers from a randomly chosen python file of the subject software package, then ranks results using a popularity index and requires to inspect only the top result. In our experiments, this method is effective at finding the true origin of a project with a recall of 0.9 and precision of 0.77. (10.1007/s10664-023-10317-8)
  DOI : 10.1007/s10664-023-10317-8
• Correlated twin-photon generation in a silicon nitride loaded thin film PPLN waveguide
  
  • Henry Antoine
  • Barral David
  • Zaquine Isabelle
  • Boes Andreas
  • Mitchell Arnan
  • Belabas Nadia
  • Bencheikh Kamel
  Optics Express, Optical Society of America - OSA Publishing, 2023, 31 (5), pp.7277. Photon-pair sources based on thin film lithium niobate on insulator technology have a great potential for integrated optical quantum information processing. We report on such a source of correlated twin-photon pairs generated by spontaneous parametric down conversion in a silicon nitride (SiN) rib loaded thin film periodically poled lithium niobate (LN) waveguide. The generated correlated photon pairs have a wavelength centred at 1560 nm compatible with present telecom infrastructure, a large bandwidth (21 THz) and a brightness of ∼2.5 × 10 5 pairs/s/mW/GHz. Using the Hanbury Brown and Twiss effect, we have also shown heralded single photon emission, achieving an autocorrelation gH(2)(0) ≃ 0.04 . (10.1364/OE.479658)
  DOI : 10.1364/OE.479658
• Localization in 1D non-parametric latent space models from pairwise affinities
  
  • Giraud Christophe
  • Issartel Yann
  • Verzelen Nicolas
  Electronic Journal of Statistics, Shaker Heights, OH : Institute of Mathematical Statistics, 2023, 17 (1), pp.1587-1662. We consider the problem of estimating latent positions in a one-dimensional torus from pairwise affinities. The observed affinity between a pair of items is modeled as a noisy observation of a function f (x*i , x*j) of the latent positions x*i, x*j of the two items on the torus. The affinity func-tion f is unknown, and it is only assumed to fulfill some shape constraints ensuring that f(x, y) is large when the distance between x and y is small, and vice-versa. This non-parametric modeling offers a good flexibility to fit data. We introduce an estimation procedure that provably localizes all the latent positions with a maximum error of the order of log(n)/n, with high-probability. This rate is proven to be minimax optimal. A computa-tionally efficient variant of the procedure is also analyzed under some more restrictive assumptions. Our general results can be instantiated to the prob-lem of statistical seriation, leading to new bounds for the maximum error in the ordering. (10.1214/23-ejs2134)
  DOI : 10.1214/23-ejs2134
• Nonatomic Non-Cooperative Neighbourhood Balancing Games
  
  • Auger David
  • Cohen Johanne
  • Lobstein Antoine
  , 2023. We introduce a game where players selfishly choose a resource and endure a cost depending on the number of players choosing nearby resources. We model the influences among resources by a weighted graph, directed or not. These games are generalizations of well-known games like Wardrop and congestion games. We study the conditions of equilibria existence and their efficiency if they exist. We conclude with studies of games whose influences among resources can be modelled by simple graphs. (10.48550/arXiv.2303.08507)
  DOI : 10.48550/arXiv.2303.08507
• Next generation of Bluetooth and Wi-Fi networks
  
  • Lim Keun-Woo
  , 2023.
• PROCÉDÉ D'ÉVALUATION DE L'ÉTAT RELATIF D'UN MOTEUR D'AÉRONEF
  
  • Pineau Edouard
  • Razakarivony Sébastien
  • Bonald Thomas
  , 2023.
• Affine invariant integrated rank-weighted statistical depth: properties and finite sample analysis
  
  • Clémençon Stephan
  • Mozharovskyi Pavlo
  • Staerman Guillaume
  Electronic Journal of Statistics, Shaker Heights, OH : Institute of Mathematical Statistics, 2023, 17 (2), pp.3854 - 3892. Because it determines a center-outward ordering of observations in Rd with d≥2, the concept of statistical depth permits to define quantiles and ranks for multivariate data and use them for various statistical tasks (e.g. inference, hypothesis testing). Whereas many depth functions have been proposed ad-hoc in the literature since the seminal contribution of [50], not all of them possess the properties desirable to emulate the notion of quantile function for univariate probability distributions. In this paper, we propose an extension of the integrated rank-weighted statistical depth (IRW depth in abbreviated form) originally introduced in [40], modified in order to satisfy the property of affine invariance, fulfilling thus all the four key axioms listed in the nomenclature elaborated by [59]. The variant we propose, referred to as the affine invariant IRW depth (AI-IRW in short), involves the precision matrix of the (supposedly square integrable) d-dimensional random vector X under study, in order to take into account the directions along which X is most variable to assign a depth value to any point x∈Rd. The accuracy of the sampling version of the AI-IRW depth is investigated from a non-asymptotic perspective. Namely, a concentration result for the statistical counterpart of the AI-IRW depth is proved. Beyond the theoretical analysis carried out, applications to anomaly detection are considered and numerical results are displayed, providing strong empirical evidence of the relevance of the depth function we propose here. (10.1214/23-EJS2189)
  DOI : 10.1214/23-EJS2189
• Online Matching in Geometric Random Graphs
  
  • Sentenac Flore
  • Noiry Nathan
  • Lerasle Matthieu
  • Ménard Laurent
  • Perchet Vianney
  , 2023. We investigate online maximum cardinality matching, a central problem in ad allocation. In this problem, users are revealed sequentially, and each new user can be paired with any previously unmatched campaign that it is compatible with. Despite the limited theoretical guarantees, the greedy algorithm, which matches incoming users with any available campaign, exhibits outstanding performance in practice. Some theoretical support for this practical success has been established in specific classes of graphs, where the connections between different vertices lack strong correlations-an assumption not always valid in real-world situations. To bridge this gap, we focus on the following model: both users and campaigns are represented as points uniformly distributed in the interval [0, 1], and a user is eligible to be paired with a campaign if they are "similar enough," meaning the distance between their respective points is less than c/N , where c > 0 is a model parameter. As a benchmark, we determine the size of the optimal offline matching in these bipartite random geometric graphs. We achieve this by introducing an algorithm that constructs the optimal matching and analyzing it. We then turn to the online setting and investigate the number of matches made by the online algorithm CLOSEST, which pairs incoming points with their nearest available neighbors in a greedy manner. We demonstrate that the algorithm's performance can be compared to its fluid limit, which is completely characterized as the solution to a specific partial differential equation (PDE). From this PDE solution, we can compute the competitive ratio of CLOSEST, and our computations reveal that it remains significantly better than its worst-case guarantee. This model turns out to be closely related to the online minimum cost matching problem, and we can extend the results obtained here to refine certain findings in that area of research. Specifically, we determine the exact asymptotic cost of CLOSEST in the ϵ-excess regime, providing a more accurate estimate than the previously known loose upper bound.
• Manifold Learning via Linear Tangent Space Alignment (LTSA) for Accelerated Dynamic MRI With Sparse Sampling
  
  • Djebra Yanis
  • Marin Thibault
  • Han Paul K
  • Bloch Isabelle
  • Fakhri Georges El
  • Ma Chao
  IEEE Transactions on Medical Imaging, Institute of Electrical and Electronics Engineers, 2023, 42 (1), pp.158-169. The spatial resolution and temporal framerate of dynamic magnetic resonance imaging (MRI) can be improved by reconstructing images from sparsely sampled k-space data with mathematical modeling of the underlying spatiotemporal signals. These models include sparsity models, linear subspace models, and non-linear manifold models. This work presents a novel linear tangent space alignment (LTSA) model-based framework that exploits the intrinsic low-dimensional manifold structure of dynamic images for accelerated dynamic MRI. The performance of the proposed method was evaluated and compared to stateof-the-art methods using numerical simulation studies as well as 2D and 3D in vivo cardiac imaging experiments. The proposed method achieved the best performance in image reconstruction among all the compared methods. The proposed method could prove useful for accelerating many MRI applications, including dynamic MRI, multi-parametric MRI, and MR spectroscopic imaging. (10.1109/TMI.2022.3207774)
  DOI : 10.1109/TMI.2022.3207774
• Proceedings of The Semantic Web: ESWC 2023 Satellite Events
  
  • Pesquita Catia
  • Skaf-Molli Hala
  • Efthymiou Vasilis
  • Kirrane Sabrina
  • Ngonga Axel
  • Collarana Diego
  • Cerqueira Renato
  • Alam Mehwish
  • Trojahn Cassia
  • Hertling Sven
  , 2023, 13998. (10.1007/978-3-031-43458-7)
  DOI : 10.1007/978-3-031-43458-7
• Dynamic Autoencoders Against Adversarial Attacks
  
  • Chabanne Hervé
  • Despiegel Vincent
  • Gentric Stéphane
  • Guiga Linda
  Procedia Computer Science, Elsevier, 2023, 220, pp.782-787. Neural Networks are the target of numerous adversarial attacks. In those, the adversary perturbs a model's input with a noise that is small, but large enough to fool the model. In this article, we propose to dynamically add autoencoders from a pretrained set to a base model as a countermeasure to such attacks. This doing, we modify the underlying labels regions of the model to be protected, letting the adversary unable to craft relevant adversarial perturbations. Our experiments confirm the efficiency of our protection when the pretrained set has enough elements. (10.1016/j.procs.2023.03.104)
  DOI : 10.1016/j.procs.2023.03.104
• Improving Causal Learning Scalability and Performance using Aggregates and Interventions
  
  • Fadiga Kanvaly
  • Houzé Etienne
  • Diaconescu Ada
  • Dessalles Jean-Louis
  ACM Transactions on Autonomous and Adaptive Systems, Association for Computing Machinery (ACM), 2023. Smart homes are Cyber-Physical Systems (CPS) where multiple devices and controllers cooperate to achieve high-level goals. Causal knowledge on relations between system entities is essential for enabling system self-adaption to dynamic changes. As house configurations are diverse, this knowledge is difficult to obtain. In previous work, we proposed to generate Causal Bayesian Networks (CBN) as follows. Starting with considering all possible relations, we progressively discarded non-correlated variables. Next, we identified causal relations from the remaining correlations by employing “ do-operations ”. The obtained CBN could then be employed for causal inference. The main challenges of this approach included: “non-doable variables” and limited scalability. To address these issues, we propose three extensions: i) early pruning weakly correlated relations to reduce the number of required do-operations; ii) introducing aggregate variables that summarize relations between weakly-coupled sub-systems; iii) applying the method a second time to perform indirect do interventions and handle non-doable relations. We illustrate and evaluate the efficiency of these contributions via examples from the smart home and power grid domain. Our proposal leads to a decrease in the number of operations required to learn the CBN and in an increased accuracy of the learned CBN, paving the way towards applications in large CPS. (10.1145/3607872)
  DOI : 10.1145/3607872
• Some Complexity Considerations on the Uniqueness of Graph Colouring
  
  • Hudry Olivier
  • Lobstein Antoine
  WSEAS Transactions on Mathematics, World Scientific and Engineering Academy and Society (WSEAS), 2023, 22, pp.art. #54, 483-493. For some well-known NP-complete problems, linked to the satisfiability of Boolean formulas and the colourability of a graph, we study the variation which consists in asking about the uniqueness of a solution.In particular, we show that five decision problems, Unique Satisfiability (U-SAT), Unique k-Satisfiability (U-k-SAT), k ≥ 3, Unique One-in-Three Satisfiability (U-1-3-SAT), Unique k-Colouring (U-k-COL), k ≥ 3, and Unique Colouring (U-COL), have equivalent complexities, up to polynomials —when dealing with colourings, we forbid permutations ofcolours. As a consequence, all are NP-hard and belong to the class DP. We also consider the problems U-2-SAT, U-2-COL and Unique Optimal Colouring (U-OCOL). (10.37394/23206.2023.22.54)
  DOI : 10.37394/23206.2023.22.54
• Towards a Development Process for Multi-CPU Distributed Synchronous Software Applications
  
  • Lubat Eric
  • Jenn Eric
  • Blouin Dominique
  • Kaufmann Marc
  Psychology in Spain, Colegio Oficial de Psicólogos, 2023, pp.549-558. (10.1109/MODELS-C59198.2023.00092)
  DOI : 10.1109/MODELS-C59198.2023.00092
• Test your samples jointly: Pseudo-reference for image quality evaluation
  
  • Tworski Marcelin
  • Lathuilière Stéphane
  , 2023. In this paper, we address the well-known image quality assessment problem but in contrast from existing approaches that predict image quality independently for every images, we propose to jointly model different images depicting the same content to improve the precision of quality estimation. This proposal is motivated by the idea that multiple distorted images can provide information to disambiguate image features related to content and quality. To this aim, we combine the feature representations from the different images to estimate a pseudo-reference that we use to enhance score prediction. Our experiments show that at test-time, our method successfully combines the features from multiple images depicting the same new content, improving estimation quality.