No description available.

The integration of large language models (LLMs) with third party applications has allowed for LLMs to retrieve information from up-to-date or specialized resources. Although this integration offers numerous advantages, it also introduces the risk of indirect prompt injection attacks. In such scenarios, an attacker embeds malicious instructions within the retrieved third party data, which when processed by the LLM, can generate harmful and untruthful outputs for an unsuspecting user. Although previous works have explored how these attacks manifest, there is no benchmarking framework to evaluate indirect prompt injection attacks and defenses at scale, limiting progress in this area. To address this gap, we introduce InjectBench, a framework that empowers the community to create and evaluate custom indirect prompt injection attack samples. Our study demonstrate that InjectBench has the capabilities to produce high quality attack samples that align with specific attack goals, and that our LLM evaluation method aligns with human judgement. Using InjectBench, we investigate the effects of different components of an attack sample on four LLM backends, and subsequently use this newly created dataset to do preliminary testing on defenses against indirect prompt injections. Experiment results suggest that while more capable models are susceptible to attacks, they are better equipped at utilizing defense strategies. To summarize, our work helps the research community to systematically evaluate features of attack samples and defenses by introducing a dataset creation and evaluation framework. / Master of Science / Large language models (LLMs), such as ChatGPT, are now able to retrieve up-to-date information from online resources like Google Flights or Wikipedia. This ultimately allows the LLM to utilize current information to generate truthful, helpful and accurate responses. Despite the numerous advantages, it also exposes a user to a new vector of attacks known as indirect prompt injections. In this attack, an attacker will write a instruction onto an online resource that an LLM will process when retrieved from the online resource. The primary aim of the attacker is to instruct the LLM to say something it is not supposed to, and thus may manifest as a blatant lie or misinformation given to the user. Prior works have studied and showcased the harmfulness of this attack, however not many works have tried to understand which LLMs are more vulnerable to indirect prompt injection attacks and how we may defend from them. We believe that this is mainly due to the non-availability of a benchmarking dataset which allows us to test LLMs and new defenses. To address this gap, we introduce InjectBench, a methodology that allows the automated creation of these benchmarking datasets, and the evaluation of LLMs and defenses. We show that InjectBench can produce a high quality dataset that we can customize to specific attack goals, and that our evaluation process is accurate and agrees with human judgement. Using the benchmarking dataset created from InjectBench, we evaluate four LLMs and investigate defenses for indirect prompt injection attacks.

Dans son dernier roman achevé intitulé Persuasion, Jane Austen se sert fréquemment du style indirect afin de représenter la pensée du personnage principal, Anne Elliot. Ce roman fut traduit en français pour la première fois en 1821 par Isabelle de Montolieu, publié à Paris sous le titre de La Famille Elliot, ou l’ancienne inclination. Qu’est-ce que le style indirect libre devient dans la narration de la première traduction française de Persuasion ? La formulation de cette question, que nous envisageons à partir d’un corpus de textes théoriques, fait suite à un certain nombre de travaux consacrés aux relations entre la traduction et le discours rapporté qui ont mis en évidence le rôle du style indirect libre dans la traduction de Montolieu. On propose pour la première fois l’application des concepts narratologiques à l’analyse de cette traduction : notre étude s’appuie ainsi sur des notions opératoires susceptibles de saisir sa singularité narrative et le rapport entre le discours rapporté et la traduction. Le troisième chapitre de notre étude débute sur l’analyse de notre traduction. Grâce aux concepts issus de la narratologie surtout à tendance « énonciative », nous avons pu regarder au-delà de la phrase pour finir par remarquer que le style indirect libre est surtout très répandu dans la narration de La Famille Elliot. Nous souhaiterions ici combler une lacune en consacrant la présente étude à un phénomène souvent jugé « extraordinaire ». Pour autant, le présent ouvrage ne doit pas être considéré uniquement comme une étude spécialisée, car il a aussi l’ambition de contribuer à l’étude du discours rapporté au sein du texte traduit en général. / In Persuasion, Jane Austen uses this technique to present Anne Elliot’s consciousness. Persuasion, Austen’s posthumously published “late” novel is first translated by Isabelle de Montolieu as La Famille Elliot, ou l’ancienne inclination, Paris, 1821. This thesis analyses the translation of FID from Persuasion to La Famille Elliot. How does Montolieu handle this technique? In chapter 1 we point out that the main reason Montolieu’s use of FID in La Famille Elliot has been neglected for so long has far less to do with Austen’s fortunes in France than with an obsession with lexical and semantic equivalence within translation studies. One of the main purposes of this study is to extend the vision of translation studies beyond the level of the sentence. We think that we achieve this by setting out to document the existence of FID in the target text narration. I argue in chapter 2 that it is impossible to comment on the narrator or FID in La Famille Elliot with any precision without first analysing definitions of these abstractions within narratology. These analytical concepts may then, only then, be applied meaningfully to the target text. Ultimately, this is what this present study does in so far as it is a target-oriented translation study that draws on key concepts from the field of narratology. A re-evaluation of the target text narration within the conceptual framework of narratology reveals extensive use of FID. In chapters 3 and 4, our analysis demonstrates that sophisticated use of FID is frequently in evidence in the target text narration. In chapter 3, we analyse several passages of FID that often function to represent the complex life of the heroine’s mind as she converses with herself. In chapter 4, we analyse numerous passages of FID that seamlessly wed the narration in La Famille Elliot to the heroine’s point of view (PDV), demonstrating that the narration achieves this focus on the heroine’s consciousness through syntactically unmarked fragments of FID thought report.

The conventional approach for testing RF circuits is specification-based testing, which involves verifying sequentially all specification requirements that are promised in the data sheet. This approach is a long-time effective testapproach but nowadays suffers from significant drawbacks.First, it requires generation and capture of test signals at the DUT operating frequency. As the operational frequencies of DUT are increasing, it becomes difficult to manage signal generation and capture using ATE. As a consequence, there is a need of expensive and specialized equipment. In addition,as conventional tests target several parameters, there is a need of several data captures and multiple test configurations. As a consequence, by adding settling time between each test and test application time, the whole test time becomes very long, and the test board very complex. Another challenge regarding RF circuit testing is wafer-level testing. Indeed, the implementation of specification-based tests at wafer level is extremely difficult due to probing issues and high parasitic effects on the test interface.Moreover, multi-site testing is usually not an option due to the small count of available RF test resources, which decreases test throughput. Hence, the current practice is often to verify the device specifications only after packaging.The problem with this solution is that defective dies are identified late in the manufacturing flow, which leads to packaging loss and decreases the global yield of the process.In order to reduce production costs, there is therefore a need to develop test solutions applicable at wafer level, so that faulty circuits can be removed very early in the production flow. This is particularly important for dies designed to be integrated in Systems-In-Package (SIP).In this context, a promising solution is to develop indirect test methods. Basically, it consists in using DUT signatures to non-conventional stimuli to predict the result of conventional tests. The underlying idea is to learn during an initial phase the unknown dependency between simple measurements and conventional tests. This dependency can then be modeled through regression functions. During the testing phase, only the indirect measurements are performed and specifications are predicted using the regression model built in the learning phase.Our work has been focused on two main directions. First, we have explored the implementation of the alternate test method based on DC measurements for RF circuits and we have proposed a methodology to select the most appropriateset of DC parameters. Results from two test vehicles (a LNA using electrical simulations and a PA using real production data) indicate that the proposed methodology allows precise estimation of the DUT performances while minimizing the number of DC measurements to be carried out.Second, we have proposed a novel implementation of the alternate test strategy in order to improve confidence in alternate test predictions and to overcome the effect of limited training set sizes. The idea is to exploit model redundancy in order to identify, during the production testing phase, devices with suspect predictions; these devices are then are removed from the alternate test tierand directed to a second tier where further testing may apply. / The conventional approach for testing RF circuits is specification-based testing, which involves verifying sequentially all specification requirements that are promised in the data sheet. This approach is a long-time effective testapproach but nowadays suffers from significant drawbacks.First, it requires generation and capture of test signals at the DUT operating frequency. As the operational frequencies of DUT are increasing, it becomes difficult to manage signal generation and capture using ATE. As a consequence, there is a need of expensive and specialized equipment. In addition,as conventional tests target several parameters, there is a need of several data captures and multiple test configurations. As a consequence, by adding settling time between each test and test application time, the whole test time becomes very long, and the test board very complex.Another challenge regarding RF circuit testing is wafer-level testing. Indeed, the implementation of specification-based tests at wafer level is extremely difficult due to probing issues and high parasitic effects on the test interface.Moreover, multi-site testing is usually not an option due to the small count of available RF test resources, which decreases test throughput. Hence, the current practice is often to verify the device specifications only after packaging.The problem with this solution is that defective dies are identified late in the manufacturing flow, which leads to packaging loss and decreases the global yield of the process.In order to reduce production costs, there is therefore a need to develop test solutions applicable at wafer level, so that faulty circuits can be removed very early in the production flow. This is particularly important for dies designed to be integrated in Systems-In-Package (SIP).In this context, a promising solution is to develop indirect test methods. Basically, it consists in using DUT signatures to non-conventional stimuli to predict the result of conventional tests. The underlying idea is to learn during an initial phase the unknown dependency between simple measurements and conventional tests. This dependency can then be modeled through regression functions. During the testing phase, only the indirect measurements are performed and specifications are predicted using the regression model built in the learning phase.Our work has been focused on two main directions. First, we have explored the implementation of the alternate test method based on DC measurements for RF circuits and we have proposed a methodology to select the most appropriateset of DC parameters. Results from two test vehicles (a LNA using electrical simulations and a PA using real production data) indicate that the proposed methodology allows precise estimation of the DUT performances while minimizing the number of DC measurements to be carried out.Second, we have proposed a novel implementation of the alternate test strategy in order to improve confidence in alternate test predictions and to overcome the effect of limited training set sizes. The idea is to exploit model redundancy in order to identify, during the production testing phase, devices with suspect predictions; these devices are then are removed from the alternate test tierand directed to a second tier where further testing may apply.

Être en mesure de vérifier si un circuit intégré est fonctionnel après fabrication peut s'avérer très difficile. Dans le cas des circuits analogiques et Radio Fréquence (RF) les procédures et les équipements de test nécessaires ont un impact majeur sur le prix de revient des circuits. Une approche intéressante pour réduire l'impact du coût du test consiste à mesurer des paramètres nécessitant des ressources de test faible coût et corréler ces mesures, dites mesures indirectes, avec les spécifications à tester. On parle alors de technique de test indirect (ou test alternatif) car il n'y a pas de mesure directe des spécifications, qui nécessiterait des équipements et du temps de test importants, mais ces spécifications sont estimées à partir des mesures « faibles couts ». Même si cette approche semble attractive elle n'est viable que si nous sommes en mesure d'établir une précision suffisante de l'estimation des performances et que cette estimation reste stable et indépendante des lots de circuits à traiter. L'objectif principal de cette thèse est de mettre en œuvre une stratégie générique permettant de proposer un flot de test indirect efficace et robuste. Pour être en mesure de construire cette stratégie nous avons amenés différentes contributions. Dans un premier temps, on a développée une nouvelle métrique dans cette thèse pour évaluer la robustesse des prédictions relaissées. Dans un deuxième temps, on a défini et analysé une stratégie pour la construction d'un model optimal. Cette dernière englobe un prétraitement de données ensuite une analyse comparative entre différentes méthodes de sélections de mesures indirectes aussi l'étude d'autres paramètres tels que la taille des combinaisons de mesures indirectes ainsi que celle de la taille de set d'apprentissage. Aussi on a proposé une stratégie pour une confidente exploration d'espace de mesures indirectes afin de construire plusieurs meilleurs modèles qu'on peut se servir par la suite pour résoudre des problèmes de confiance et d'optimisation. Les études comparatives réalisées ont été effectuées sur 2 cas d'études expérimentaux et à partir de métriques classiques et de la nouvelle métrique proposée permettant ainsi d'évaluer objectivement la robustesse de chaque solution.En fin, nous avons développé une stratégie complète mettant en œuvre des techniques de redondance de modèles de corrélation qui permettent d'améliorer grandement la robustesse et l'efficacité de la prise de décision en fonction des mesures obtenues. Cette stratégie est adaptable à n'importe quel contexte en termes de compromis entre le coût du test et le niveau de confiance et de précision attendu. / Being able to check whether an IC is functional or not after the manufacturing process is very difficult. Particularly for analog and Radio Frequency (RF) circuits, test equipment and procedures required have a major impact on the circuits cost. An interesting approach to reduce the impact of the test cost is to measure parameters requiring low cost test resources and correlate these measurements, called indirect measurements, with the targeted specifications. This is known as indirect test technique because there is no direct measurement for these specifications, which requires so expensive test equipment and an important testing time, but these specifications are estimated w.r.t "low-cost measurements". While this approach seems attractive, it is only viable if we are able to establish a sufficient accuracy for the performance estimation and if this estimation remains stable and independent from the circuits sets under test.The main goal of this thesis is to implement a robust and effective indirect test strategy for a given application and to improve test decisions based on data analysis.To be able to build this strategy, we have brought various contributions. Initially, we have defined new metric developed in this thesis to assess the reliability of the estimated performances. Secondly, we have analyzed and defined a strategy for the construction of an optimal model. This latter includes a data preprocessing followed by a comparative analysis of different methods of indirect measurement selection. Then, we have proposed a strategy for a confidant exploration of the indirect measurement space in order to build several best models that can be used later to solve trust and optimization issues. Comparative studies were performed on 2 experimental data sets by using both of the conventional and the developed metrics to evaluate the robustness of each solution in an objective way.Finally, we have developed a comprehensive strategy based on an efficient implementation of the redundancy techniques w.r.t to the build models. This strategy has greatly improved the robustness and the effectiveness of the decision plan based on the obtained measurements. This strategy is adaptable to any context in terms of compromise between the test cost, the confidence level and the expected precision.

Netiesioginiu mokesčių harmonizavimo proceso įtaka yra aktualiausia visuomenei. Harmonizavimo pasekmės įtakoja vartojimo išlaidas, lemia kainų didėjimą. Lietuvoje taikomi netiesioginių mokesčių tarifai yra derinami su Europos Sąjungoje taikomais minimaliais tarifais. Šio darbo autorius nagrinėja netiesioginių mokesčių harmonizavimo procesą Lietuvoje bei Europos Sąjungoje. Pagrindinis dėmesys skirtas harmonizavimo proceso poveikio namų ūkiams analizei. Palyginti taikomi PVM tarifai ir lengvatos valstybėse narėse. Apibendrinti apmokestinimo akcizo mokesčių pasikeitimai, įstojus į Europos Sąjungą. Išanalizuoti alui, vynui, benzinui, cigarams ir cigarilėms taikomi akcizo mokesčio tarifai valstybėse narėse. Darbo pabaigoje pateiktos išvados ir pasiūlymai. Baigiamajame darbe buvo naudojami empiriniai tyrimo metodai: kontent analizė, tradicinė, dokumentų analizė. Gauti empiriniai tyrimo duomenys apibendrinti lyginimo ir analizės metodais. / The impact of indirect tax harmonization process is the most relevant aspect in respect to the society. Consequences of the harmonization are of the great effect to consumer's expenditures and it leads to an increase in prices. The tariffs of indirect taxes applied in Lithuania are harmonized in accordance with the minimal tariffs applied in the EU. The key aspect of the thesis is addressed towards the analysis of an effect of the harmonization on households.

This thesis contains a critical review of the most important approaches to reported speech or those pertinent to it in the linguistic literature. Several problems concerning quotation parentheticals and presuppositions have been discussed and some solutions are proposed. The main issue, referred to as "the point of view", is worked out as an extension of the classical notion of transparency and opacity applied to NP, to several other elements of sentence structures. As a result, a set of interpretive rules are formulated which specify whether certain linguistic configurations promote a transparent or opaque interpretation of the sentence or part of it. Also, rules of correspondence are given which stipulate the conditions for semantic well formedness, in the case of omission of information.

La contribution relative du bruit de combustion au bruit global des moteurs aéronautiques augmente progressivement dû d'un coté à la réduction des autres sources et d'un autre à l'implémentation des nouvelles technologies de chambre de combustion pour la réduction des émissions de NOx. Deux mécanismes sont responsables de cette source de bruit : d'abord le bruit direct, dû aux ondes acoustiques générées par la flamme qui se propagent à la sortie du moteur d'avion, et ensuite le bruit indirect, généré par les ondes d'entropie quand elles sont accélérées et ralenties dans les étages de turbine. Dans ce travail, les modèles analytiques utilisés pour la propagation des ondes à travers les flux non-homogènes, y compris la génération de bruit indirect, sont révisés et étendus. Tout d'abord, le cas quasi-1D est étudié: la méthode analytique est étendue pour les fréquences non nulles et validée avec des méthodes numériques et des données expérimentales. Dans la seconde partie, la méthode analytique 2D dans le cas d'aubes de turbines compactes est étudiée et validée à l'aide de simulations numériques d'un rotor et d'un étage de turbine complète. Enfin, ces modèles sont combinés avec des simulations aux grandes échelles réactives et compressibles de chambres de combustion pour construire une approche hybride appelée CHORUS capable de prédire le bruit de combustion. / Combustion noise is increasing its relative contribution to aircraft noise, while other sources are being reduced and new low-NOx emission combustion chambers being built. Two mechanisms are responsible for this noise source: direct noise in which acoustic waves are generated by the flame and propagate to the outlet of the aero-engine, and indirect noise, where entropy waves generate noise as they are accelerated and decelerated in the turbine stages. In this work the analytical models used for the propagation of waves through non-homogeneous flows, including the generation of indirect noise, are revised and extended. In the first part, the quasi-1D case is studied, extending the analytical method to non-zero frequencies and validating the results with numerical methods and experimental data. In the second part, the 2D method for the case of compact turbine blades is studied and validated using numerical simulations of a rotating blade and of a complete turbine stage. Finally, in the third part of this thesis, these models are combined with reactive and compressible Large Eddy Simulations (LES) of combustion chambers to build a hybrid approach, named CHORUS, able to predict combustion noise.

No description available.

No description available.