Publications

2021

• SAMbA: Speech enhancement with Asynchronous ad-hoc Microphone Arrays
  
  • Furnon Nicolas
  • Serizel Romain
  • Essid Slim
  • Illina Irina
  , 2021. Speech enhancement in ad-hoc microphone arrays is often hindered by the asynchronization of the devices composing the microphone array. Asynchronization comes from sampling time offset and sampling rate offset which inevitably occur when the microphones are embedded in different hardware components. In this paper, we propose a deep neural network (DNN)-based speech enhancement solution that is suited for applications in ad-hoc microphone arrays because it is distributed and copes with asynchronization. We show that asynchronization has a limited impact on the spatial filtering and mostly affects the performance of the DNNs. Instead of resynchronising the signals, which requires costly processing steps, we use an attention mechanism which makes the DNNs, thus our whole pipeline, robust to asynchronization. We also show that the attention mechanism leads to the asynchronization parameters in an unsupervised manner. (10.48550/arXiv.2307.16582)
  DOI : 10.48550/arXiv.2307.16582
• Quantum cryptography in a hybrid security model
  
  • Vyas Nilesh
  , 2021. Extending the functionality and overcoming the performance limitation of QKD requires either quantum repeaters or new security models. Investigating the latter option, we introduce the Quantum Computational Timelock (QCT) security model, assuming that computationally secure encryption may only be broken after time much longer than the coherence time of available quantum memories. These two assumptions, namely short-term computational security and noisy quantum storage, have so far already been considered in quantum cryptography, yet only disjointly. A practical lower bound on time, for which encryption is computationally secure, can be inferred from assumed long-term security of the AES256 encryption scheme (30 years) and the value of coherence time in experimental demonstrations of storage and then retrieval of optically encoded quantum information, at single-photon level range from a few nanoseconds to microseconds. Given the large gap between the upper bound on coherence time and lower bound on computational security time of an encryption scheme, the validity of the QCT security model can be assumed with a very high confidence today and also leaves a considerable margin for its validity in the future. Using the QCT security model, we propose an explicit d-dimensional key agreement protocol that we call MUB-Quantum Computational Timelock (MUB-QCT), where a bit is encoded on a qudit state using a full set of mutually unbiased bases (MUBs) and a family of pair-wise independent permutations. Security is proved by showing that upper bound on Eve's information scales as O(1=d). We show MUB-QCT offers: high resilience to error (up to 50% for large d) with fixed hardware requirements; MDI security as security is independent of channel monitoring and does not require to trust measurement devices. We also prove the security of the MUB-QCT protocol, with multiple photons per channel use, against non-adaptive attacks, in particular, proactive MUB measurement where eve measures each copy in a different MUB followed by post-measurement decoding. We prove that the MUB-QCT protocol allows secure key distribution with input states containing up to O(d) photons which implies a significant performance boost, characterized by an O(d) multiplication of key rate and a significant increase in the reachable distance. These results illustrate the power of the QCT security model to boost the performance of quantum cryptography while keeping a clear security advantage over classical cryptography. (10.70675/71129bffza729z4b97zbdb0z152ee4a2fcab)
  DOI : 10.70675/71129bffza729z4b97zbdb0z152ee4a2fcab
• Video compression of airplane cockpit screens content
  
  • Mitrica Iulia
  , 2021. This thesis addresses the problem of encoding the video of airplane cockpits.The cockpit of modern airliners consists in one or more screens displaying the status of the plane instruments (e.g., the plane location as reported by the GPS, the fuel level as read by the sensors in the tanks, etc.,) often superimposed over natural images (e.g., navigation maps, outdoor cameras, etc.).Plane sensors are usually inaccessible due to security reasons, so recording the cockpit is often the only way to log vital plane data in the event of, e.g., an accident.Constraints on the recording storage available on-board require the cockpit video to be coded at low to very low bitrates, whereas safety reasons require the textual information to remain intelligible after decoding. In addition, constraints on the power envelope of avionic devices limit the cockpit recording subsystem complexity.Over the years, a number of schemes for coding images or videos with mixed computer-generated and natural contents have been proposed. Text and other computer generated graphics yield high-frequency components in the transformed domain. Therefore, the loss due to compression may hinder the readability of the video and thus its usefulness. For example, the recently standardized Screen Content Coding (SCC) extension of the H.265/HEVC standard includes tools designed explicitly for screen contents compression. Our experiments show however that artifacts persist at the low bitrates targeted by our application, prompting for schemes where the video is not encoded in the pixel domain.This thesis proposes methods for low complexity screen coding where text and graphical primitives are encoded in terms of their semantics rather than as blocks of pixels.At the encoder side, characters are detected and read using a convolutional neural network.Detected characters are then removed from screen via pixel inpainting, yielding a smoother residual video with fewer high frequencies. The residual video is encoded with a standard video codec and is transmitted to the receiver side together with text and graphics semantics as side information.At the decoder side, text and graphics are synthesized using the decoded semantics and superimposed over the residual video, eventually recovering the original frame. Our experiments show that an AVC/H.264 encoder retrofitted with our method has better rate-distortion performance than H.265/HEVC and approaches that of its SCC extension.If the complexity constraints allow inter-frame prediction, we also exploit the fact that co-located characters in neighbor frames are strongly correlated.Namely, the misclassified symbols are recovered using a proposed method based on low-complexity model of transitional probabilities for characters and graphics. Concerning character recognition, the error rate drops up to 18 times in the easiest cases and at least 1.5 times in the most difficult sequences despite complex occlusions.By exploiting temporal redundancy, our scheme further improves in rate-distortion terms and enables quasi-errorless character decoding. Experiments with real cockpit video footage show large rate-distortion gains for the proposed method with respect to video compression standards. (10.70675/18512bbdzfb2fz4993zaecfz8a0e94302a12)
  DOI : 10.70675/18512bbdzfb2fz4993zaecfz8a0e94302a12
• Quantum-dot lasers on silicon : nonlinear properties, dynamics, and applications
  
  • Dong Bozhang
  , 2021. Silicon photonics is promising for high-speed communication systems, short-reach optical interconnects, and quantum technologies. Direct epitaxial growth of III-V materials on silicon is also an ideal solution for the next generation of photonic integrated circuits (PICs). In this context, quantum-dots (QD) lasers with atom-like density of states are promising to serve as the on-chip laser sources, owing to their high thermal stability and strong tolerance for the defects that arise during the epitaxial growth. The purpose of this dissertation is to investigate the nonlinear properties and dynamics of QD lasers on Si for PIC applications. The first part of this thesis investigates the dynamics of epitaxial QD lasers on Si subject to external optical feedback (EOF). In the short-cavity regime, the QD laser exhibits strong robustness against parasitic reflections hence giving further insights for developing isolator-free PICs. In particular, a near-zero linewidth enhancement factor is crucial to achieve this goal. The second part is devoted to studying the static properties and dynamics of a single-frequency QD distributed feedback (DFB) laser for uncooled and isolator-free applications. The design of a temperature-controlled mismatch between the optical gain peak and the DFB wavelength contributes to improving the laser performance with the increase of temperature. The third part of this dissertation investigates the QD-based optical frequency comb (OFC). External control techniques including EOF and optical injection are used to optimize the noise properties, reduce the timing-jitter, and increase the frequency comb bandwidth. In the last part, an investigation of the optical nonlinearities of the QD laser on Si is carried out by the four-wave mixing (FWM) effect. This study demonstrates that the FWM efficiency of QD laser is more than one order of magnitude higher than that of a commercial quantum-well laser, which gives insights for developing self-mode-locked OFCs based on QD. All these results allow for a better understanding of the nonlinear dynamics of QD lasers and pave the way for developing high-performance classical and quantum PICs on Si. (10.70675/02f50e71z4324z4375za91ez03b01f03de51)
  DOI : 10.70675/02f50e71z4324z4375za91ez03b01f03de51
• Cryogenic In-MRAM Computing
  
  • Hou Yaoru
  • Ge We
  • Guo Yanan
  • Naviner Lirida
  • Wang You
  • Liu Bo
  • Yang Jun
  • Cai Hao
  , 2021, pp.1-6. In the computation storage separated von-Neumann architecture, memory-wall becomes critical due to large access latency and tremendous amount of data movement. In this work, we pursue cryogenic temperature based memory design and focus on spin-transfer-torque magnetoresistive random access memory (STT-MRAM) at 77-Kelvin (achieved with low-cost liquid nitrogen). Cryogenic compact model and related cryogenic bitcell are investigated based on 77K experiment data of magnetic tunnel junction (MTJ) and CMOS transistor. Aggressive energy reduction is obtained through in-MRAM computing architecture. A 1Kb sub-array is simulated based on above cryogenic models. Results show that cryogenic in-MRAM computing provides performance improvements of 32% on average, and concurrently reduces memory energy consumption by 19% on average. Compared with room temperature (RT) simulation results, a 70% reduction of sensing latency is realized at 0.7-V supply voltage, with the cost of 30% increased writing latency and 20% higher energy consumption. A 32.5% sensing failure probability is alleviated in the 77K cryogenic environment. The proposed 77K cryogenic design methodology can be further applied to energy constrained applications. (10.1109/NANOARCH53687.2021.9642238)
  DOI : 10.1109/NANOARCH53687.2021.9642238
• Personalized audio auto-tagging as proxy for contextual music recommendation
  
  • Ibrahim Karim Magdi Abdelfattah
  , 2021. The exponential growth of online services and user data changed how we interact with various services, and how we explore and select new products. Hence, there is a growing need for methods to recommend the appropriate items for each user. In the case of music, it is more important to recommend the right items at the right moment. It has been well documented that the context, i.e. the listening situation of the users, strongly influences their listening preferences. Hence, there has been an increasing attention towards developing recommendation systems. State-of-the-art approaches are sequence-based models aiming at predicting the tracks in the next session using available contextual information. However, these approaches lack interpretability and serve as a hit-or-miss with no room for user involvement. Additionally, few previous approaches focused on studying how the audio content relates to these situational influences, and even to a less extent making use of the audio content in providing contextual recommendations. Hence, these approaches suffer from both lack of interpretability.In this dissertation, we study the potential of using the audio content primarily to disambiguate the listening situations, providing a pathway for interpretable recommendations based on the situation.First, we study the potential listening situations that influence/change the listening preferences of the users. We developed a semi-automated approach to link between the listened tracks and the listening situation using playlist titles as a proxy. Through this approach, we were able to collect datasets of music tracks labelled with their situational use. We proceeded with studying the use of music auto-taggers to identify potential listening situations using the audio content. These studies led to the conclusion that the situational use of a track is highly user-dependent. Hence, we proceeded with extending the music-autotaggers to a user-aware model to make personalized predictions. Our studies showed that including the user in the loop significantly improves the performance of predicting the situations. This user-aware music auto-tagger enabled us to tag a given track through the audio content with potential situational use, according to a given user by leveraging their listening history.Finally, to successfully employ this approach for a recommendation task, we needed a different method to predict the potential current situations of a given user. To this end, we developed a model to predict the situation given the data transmitted from the user's device to the service, and the demographic information of the given user. Our evaluations show that the models can successfully learn to discriminate the potential situations and rank them accordingly. By combining the two model; the auto-tagger and situation predictor, we developed a framework to generate situational sessions in real-time and propose them to the user. This framework provides an alternative pathway to recommending situational sessions, aside from the primary sequential recommendation system deployed by the service, which is both interpretable and addressing the cold-start problem in terms of recommending tracks based on their content. (10.70675/60b82302z46e2z4538za84bz80fd5b342714)
  DOI : 10.70675/60b82302z46e2z4538za84bz80fd5b342714
• Security Analysis of Machine Learning-Based PUF Enrollment Protocols: A Review
  
  • Khalfaoui Sameh
  • Leneutre Jean
  • Villard Arthur
  • Gazeau Ivan
  • Ma Jingxuan
  • Urien Pascal
  Sensors, MDPI, 2021, 21 (24), pp.8415. The demand for Internet of Things services is increasing exponentially, and consequently a large number of devices are being deployed. To efficiently authenticate these objects, the use of physical unclonable functions (PUFs) has been introduced as a promising solution for the resource-constrained nature of these devices. The use of machine learning PUF models has been recently proposed to authenticate the IoT objects while reducing the storage space requirement for each device. Nonetheless, the use of a mathematically clonable PUFs requires careful design of the enrollment process. Furthermore, the secrecy of the machine learning models used for PUFs and the scenario of leakage of sensitive information to an adversary due to an insider threat within the organization have not been discussed. In this paper, we review the state-of-the-art model-based PUF enrollment protocols. We identity two architectures of enrollment protocols based on the participating entities and the building blocks that are relevant to the security of the authentication procedure. In addition, we discuss their respective weaknesses with respect to insider and outsider threats. Our work serves as a comprehensive overview of the ML PUF-based methods and provides design guidelines for future enrollment protocol designers. (10.3390/s21248415)
  DOI : 10.3390/s21248415
• Kalman Filtering for Learning with Evolving Data Streams
  
  • Ziffer Giacomo
  • Bernardo Alessio
  • Valle Emanuele Della
  • Bifet Albert
  , 2021, pp.5337--5346. Processing data streams gained much importance in recent years. Standard machine learning algorithms do not cope well with non-stationary streaming data, where decision models evolve and generate so-called concept drift. Online adaptive algorithms emerged to solve these issues. They learn incrementally and generally require explicit forgetting mechanisms to adapt to concept drift. In this paper, we propose the application of Kalman filtering to handle evolving data streams. This novel approach addresses data stream mining and concept drift management challenges from a new perspective, directly modelling a representation suitable for the data streams. First, we study a Kalman filter based learning a pproach and investigate its integration into the Naïve Bayes algorithm, namely KalmanNB. Additionally, we propose the Hoeffding Kalman Tree, a combination of the Hoeffding Tree with KalmanNB. Empirical results demonstrate that the Kalman filter based approach inherently manages concept drifts, and it adapts to the emerging concept more rapidly than the state-of-the-art algorithms. Moreover, it is an accurate and robust approach and requires less storage while still being faster. (10.1109/BIGDATA52589.2021.9671365)
  DOI : 10.1109/BIGDATA52589.2021.9671365
• Neural methods for spoken dialogue understanding
  
  • Chapuis Emile
  , 2021. Conversational AI has received a growing interest in recent years from both the research community and the industry. Products have started to emerge (e.g. Amazon's Alexa, Google's Home, Apple's Siri) but performances of such systems are still far from human-likeness communication. As an example, conversation with the aforementioned systems is often limited to basic question-response interactions. Among all the reasons why people communicate, the exchange of information and the strengthening of social bound appeared to be the main ones. In dialogue research, the two aforementioned problems are well known and addressed using dialogue act classification and emotion/sentiment recognition. Those problems are made even more challenging as they involve spoken dialogues in contrast to written text. A spoken conversation is a complex and collective activity that has a specific dynamic and structure. Thus, there is a need to adapt both natural language processing and natural language understanding techniques which have been tailored for written texts as it does not share the same characteristics. This thesis focuses on methods for spoken dialogue understanding and specifically tackles the problem of spoken dialogues classification with a particular focus on dialogue act and emotion/sentiment labels. Our contributions can be divided into two parts: in the first part, we address the problem of automatically labelling English spoken dialogues. In this part, we start by formulating this problem as a translation problem which leads us to propose a seq2seq model for dialogue act classification. Then, our second contribution focuses on a scenario relying on small annotated datasets and involves both pre-training a hierarchical transformer encoder and proposing a new benchmark for evaluation. This first part addresses the problem of spoken language classification in monolingual (i.e. English) and monomodal (i.e. text) settings. However, spoken dialogue involves phenomena such as code-switching (when a speaker switch languages within a conversation) and relies on multiple channels to communicate (e.g.} audio or visual).Hence, the second part is dedicated to two extensions of the previous contributions in two settings: multilingual and multimodal. We first address the problem of dialogue act classification when multiple languages are involved and thus, we extend the two previous contributions to a multilingual scenario. In our last contribution, we explore a multimodal scenario and focus on the representation and fusion of modalities in the scope of emotion prediction. (10.70675/1fc9d449zab9az41c9zaa43z5f3ca7ec8ba4)
  DOI : 10.70675/1fc9d449zab9az41c9zaa43z5f3ca7ec8ba4
• Challenges of Machine Learning for Data Streams in the Banking Industry
  
  • Barry Mariam
  • Bifet Albert
  • Chiky Raja
  • Montiel Jacob
  • Tran Vinh-Thuy
  , 2021, 13147, pp.106--118. Banking Information Systems continuously generate large quantities of data as inter-connected streams (transactions, events logs, time series, metrics, graphs, process, etc.). Such data streams need to be processed online to deal with critical business applications such as real-time fraud detection, network security attack prevention or predictive maintenance on information system infrastructure. Many algorithms have been proposed for data stream learning, however, most of them do not deal with the important challenges and constraints imposed by real-world applications. In particular, when we need to train models incrementally from heterogeneous data mining and deployment them within complex big data architecture. Based on banking applications and lessons learned in production environments of BNP Paribas - a major international banking group and leader in the Eurozone - we identified the most important current challenges for mining IT data streams. Our goal is to highlight the key challenges faced by data scientists and data engineers within complex industry settings for building or deploying models for real word streaming applications. We provide future research directions on Stream Learning that will accelerate the adoption of online learning models for solving real-word problems. Therefore bridging the gap between research and industry communities. Finally, we provide some recommendations to tackle some of these challenges. (10.1007/978-3-030-93620-4\_9)
  DOI : 10.1007/978-3-030-93620-4\_9
• Nouvelles techniques de compression pour le codage vidéo prochaine-génération
  
  • Nasrallah Anthony
  , 2021. Video content now occupies about 82% of global internet traffic. This large percentage is due to the revolution in video content consumption. On the other hand, the market is increasingly demanding videos with higher resolutions and qualities. This causes a significant increase in the amount of data to be transmitted. Hence the need to develop video coding algorithms even more efficient than existing ones to limit the increase in the rate of data transmission and ensure a better quality of service. In addition, the impressive consumption of multimedia content in electronic products has an ecological impact. Therefore, finding a compromise between the complexity of algorithms and the efficiency of implementations is a new challenge. As a result, a collaborative team was created with the aim of developing a new video coding standard, Versatile Video Coding – VVC/H.266. Although VVC was able to achieve a more than 40% reduction in throughput compared to HEVC, this does not mean at all that there is no longer a need to further improve coding efficiency. In addition, VVC adds remarkable complexity compared to HEVC. This thesis responds to these problems by proposing three new encoding methods. The contributions of this research are divided into two main axes. The first axis is to propose and implement new compression tools in the new standard, capable of generating additional coding gains. Two methods have been proposed for this first axis. These two methods rely on the derivation of prediction information at the decoder side. This is because increasing encoder choices can improve the accuracy of predictions and yield less energy residue, leading to a reduction in bit rate. Nevertheless, more prediction modes involve more signaling to be sent into the binary stream to inform the decoder of the choices that have been made at the encoder. The gains mentioned above are therefore more than offset by the added signaling. If the prediction information has been derived from the decoder, the latter is no longer passive, but becomes active hence the concept of intelligent decoder. Thus, it will be useless to signal the information, hence a gain in signalization. Each of the two methods offers a different intelligent technique than the other to predict information at the decoder level. The first technique constructs a histogram of gradients to deduce different intra-prediction modes that can then be combined by means of prediction fusion, to obtain the final intra-prediction for a given block. This fusion property makes it possible to more accurately predict areas with complex textures, which, in conventional coding schemes, would rather require partitioning and/or finer transmission of high-energy residues. The second technique gives VVC the ability to switch between different interpolation filters of the inter prediction. The deduction of the optimal filter selected by the encoder is achieved through convolutional neural networks. The second axis, unlike the first, does not seek to add a contribution to the VVC algorithm. This axis rather aims to build an optimized use of the already existing algorithm. The ultimate goal is to find the best possible compromise between the compression efficiency delivered and the complexity imposed by VVC tools. Thus, an optimization system is designed to determine an effective technique for activating the new coding tools. The determination of these tools can be done either using artificial neural networks or without any artificial intelligence technique. (10.70675/53731967z7c74z47c2za238z4d87ca6b3a03)
  DOI : 10.70675/53731967z7c74z47c2za238z4d87ca6b3a03
• Conceptual Security Architecture Interim Report, QSAFE: Detailed Study for the European Quantum Communication Infrastructure
  
  • Alleaume Romain
  , 2021.
• Efficient Deep Learning of Nonlinear Fiber-Optic Communications Using a Convolutional Recurrent Neural Network
  
  • Shahkarami Abtin
  • Yousefi Mansoor
  • Jaouën Yves
  , 2022. Nonlinear channel impairments are a major obstacle in fiber-optic communication systems. To facilitate a higher data rate in these systems, the complexity of the underlying digital signal processing algorithms to compensate for these impairments must be reduced. Deep learning-based methods have proven successful in this area. However, the concept of computational complexity remains an open problem. In this paper, a lowcomplexity convolutional recurrent neural network (CNN+RNN) is considered for deep learning of the long-haul optical fiber communication systems where the channel is governed by the nonlinear Schrödinger equation. This approach reduces the computational complexity via balancing the computational load by capturing short-temporal distance features using strided convolution layers with ReLU activation, and the long-distance features using a many-to-one recurrent layer. We demonstrate that for a 16-QAM 100 G symbol/s system over 2000 km optical-link of 20 spans, the proposed approach achieves the bit-error-rate of the digital back-propagation (DBP) with substantially fewer floating-point operations (FLOPs) than the recently-proposed learned DBP, as well as the non-model-driven deep learningbased equalization methods using end-to-end MLP, CNN, RNN, and bi-RNN models. Index Terms-Fiber-optic communications, deep learning, nonlinear channel impairments, convolutional recurrent neural networks. (10.1109/icmla52953.2021.00112)
  DOI : 10.1109/icmla52953.2021.00112
• Formal Evaluation and Construction of Glitch-resistant Masked Functions
  
  • Takarabt Sofiane
  • Guilley Sylvain
  • Souissi Youssef
  • Karray Khaled
  • Sauvage Laurent
  • Mathieu Yves
  , 2021.
• Neuro-steered music source separation
  
  • Cantisani Giorgia
  , 2021. In this PhD thesis, we address the challenge of integrating Brain-Computer Interfaces (BCI) and music technologies on the specific application of music source separation, which is the task of isolating individual sound sources that are mixed in the audio recording of a musical piece. This problem has been investigated for decades, but never considering BCI as a possible way to guide and inform separation systems. Specifically, we explored how the neural activity characterized by electroencephalographic signals (EEG) reflects information about the attended instrument and how we can use it to inform a source separation system.First, we studied the problem of EEG-based auditory attention decoding of a target instrument in polyphonic music, showing that the EEG tracks musically relevant features which are highly correlated with the time-frequency representation of the attended source and only weakly correlated with the unattended one. Second, we leveraged this ``contrast'' to inform an unsupervised source separation model based on a novel non-negative matrix factorisation (NMF) variant, named contrastive-NMF (C-NMF) and automatically separate the attended source.Unsupervised NMF represents a powerful approach in such applications with no or limited amounts of training data as when neural recording is involved. Indeed, the available music-related EEG datasets are still costly and time-consuming to acquire, precluding the possibility of tackling the problem with fully supervised deep learning approaches. Thus, in the last part of the thesis, we explored alternative learning strategies to alleviate this problem. Specifically, we propose to adapt a state-of-the-art music source separation model to a specific mixture using the time activations of the sources derived from the user's neural activity. This paradigm can be referred to as one-shot adaptation, as it acts on the target song instance only.We conducted an extensive evaluation of both the proposed system on the MAD-EEG dataset which was specifically assembled for this study obtaining encouraging results, especially in difficult cases where non-informed models struggle. (10.70675/0fd7a113zc08fz49bbza24dz94a0834ee136)
  DOI : 10.70675/0fd7a113zc08fz49bbza24dz94a0834ee136
• Accountability and reconfiguration: self-healing lattice agreement
  
  • Freitas Luciano
  • Kuznetsov Petr
  • Rieutord Thibault
  • Tucci Sara
  , 2022, pp.25. An accountable distributed system provides means to detect deviations of system components from their expected behavior. It is natural to complement fault detection with a reconfiguration mechanism, so that the system could heal itself, by replacing malfunctioning parts with new ones. In this paper, we describe a framework that can be used to implement a large class of accountable and reconfigurable replicated services. We build atop the fundamental lattice agreement abstraction lying at the core of storage systems and cryptocurrencies. Our asynchronous implementation of accountable lattice agreement ensures that every violation of consistency is followed by an undeniable evidence of misbehavior of a faulty replica. The system can then be seamlessly reconfigured by evicting faulty replicas, adding new ones and merging inconsistent states. We believe that this paper opens a direction towards asynchronous “self-healing” systems that combine accountability and reconfiguration (10.4230/LIPIcs.OPODIS.2021.25)
  DOI : 10.4230/LIPIcs.OPODIS.2021.25
• Practical Privacy-Preserving Face Identification based on Function-Hiding Functional Encryption
  
  • Ibarrondo Alberto
  • Chabanne Hervé
  • Önen Melek
  , 2021, 13099, pp.63-71. Leveraging on function-hiding Functional Encryption (FE) and inner-product-based matching, this work presents a practical privacypreserving face identification system with two key novelties: switching functionalities of encryption and key generation algorithms of FE to optimize matching latency while maintaining its security guarantees, and identifying output leakage to later formalize two new attacks based on it with appropriate countermeasures. We validate our scheme in a realistic face matching scenario, attesting its applicability to pseudo real-time one-use face identification scenarios like passenger identification. (10.1007/978-3-030-92548-2_4)
  DOI : 10.1007/978-3-030-92548-2_4
• Metasuperfície de Huygens reconfigurável para realização de um beamsteering independente na banda 5G sem impacto nas bandas 4G e Wi-Fi 2,4/5/6E
  
  • Gonçalves Licursi de Mello Rafael
  • Lepage Anne Claire
  • Begaud Xavier
  , 2021.
• What can information guess ? : Towards information leakage quantification in side-channel analysis
  
  • Cheng Wei
  , 2021. Cryptographic algorithms are nowadays prevalent in establishing secure connectivity in our digital society. Such computations handle sensitive information like encryption keys, which are usually very exposed during manipulation, resulting in a huge threat to the security of the sensitive information concealed in cryptographic components. In the field of embedded systems security, side-channel analysis is one of the most powerful techniques against cryptographic implementations. The main subject of this thesis is the measurable side-channel security of cryptographic implementations, particularly in the presence of random masking. Overall, this thesis consists of two topics. One is the leakage quantification of the most general form of masking equipped with the linear codes, so-called code-based masking; the other one is exploration of applying more generic information measures in a context of side-channel analysis. Two topics are inherently connected to each other in assessing and enhancing the practical security of cryptographic implementations .Regarding the former, we propose a unified coding-theoretic framework for measuring the information leakage in code-based masking. Specifically, our framework builds formal connections between coding properties and leakage metrics in side-channel analysis. Those formal connections enable us to push forward the quantitative evaluation on how the linear codes can affect the concrete security of all code-based masking schemes. Moreover, relying on our framework, we consolidate code-based masking by providing the optimal linear codes in the sense of maximizing the side-channel resistance of the corresponding masking scheme. Our framework is finally verified by attack-based evaluation, where the attacks utilize maximum-likelihood based distinguishers and are therefore optimal. Regarding the latter, we present a full spectrum of application of alpha-information, a generalization of (Shannon) mutual information, for assessing side-channel security. In this thesis, we propose to utilize a more general information-theoretic measure, namely alpha-information (alpha-information) of order alpha. The new measure also gives the upper bound on success rate and the lower bound on the number of measurements. More importantly, with proper choices of alpha, alpha-information provides very tight bounds, in particular, when alpha approaches to positive infinity, the bounds will be exact. As a matter of fact, maximum-likelihood based distinguishers will converge to the bounds. Therefore, we demonstrate how the two world, information-theoretic measures (bounds) and maximum-likelihood based side-channel attacks, are seamlessly connected in side-channel analysis .In summary, our study in this thesis pushes forward the evaluation and consolidation of side-channel security of cryptographic implementations. From a protection perspective, we provide a best-practice guideline for the application of code-based masking. From an evaluation perspective, the application of alpha-information enables practical evaluators and designers to have a more accurate (or even exact) estimation of concrete side-channel security level of their cryptographic chips. (10.70675/264cf853z73adz419fz8ef6z12f6ffa85f33)
  DOI : 10.70675/264cf853z73adz419fz8ef6z12f6ffa85f33
• Parasite: Mitigating Physical Side-Channel Attacks Against Neural Networks
  
  • Chabanne Hervé
  • Danger Jean-Luc
  • Guiga Linda
  • Kühne Ulrich
  , 2022, 13162, pp.148-167. Neural Networks (NNs) are now the target of various side-channel attacks whose aim is to recover the model’s parameters and/or architecture. We focus our work on EM side-channel attacks for parameter extraction. We propose a novel approach to countering such side-channel attacks, based on the method introduced by Chabanne et al. in 2021, where parasitic convolutional models are dynamically applied to the input of the victim model. We validate this new idea in the side-channel field by simulation (10.1007/978-3-030-95085-9_8)
  DOI : 10.1007/978-3-030-95085-9_8
• Informed audio source separation with deep learning in limited data settings
  
  • Schulze-Forster Kilian
  , 2021. Audio source separation is the task of estimating the individual signals of several sound sources when only their mixture can be observed. State-of-the-art performance for musical mixtures is achieved by Deep Neural Networks (DNN) trained in a supervised way. They require large and diverse datasets of mixtures along with the target source signals in isolation. However, it is difficult and costly to obtain such datasets because music recordings are subject to copyright restrictions and isolated instrument recordings may not always exist.In this dissertation, we explore the usage of additional information for deep learning based source separation in order to overcome data limitations.First, we focus on a supervised setting with only a small amount of available training data. We investigate to which extent singing voice separation can be improved when it is informed by lyrics transcripts. To this end, a novel deep learning model for informed source separation is proposed. It aligns text and audio during the separation using a novel monotonic attention mechanism. The lyrics alignment performance is competitive with state-of-the-art methods while a smaller amount of training data is used. We find that exploiting aligned phonemes can improve singing voice separation, but precise alignments and accurate transcripts are required.Finally, we consider a scenario where only mixtures but no isolated source signals are available for training. We propose a novel unsupervised deep learning approach to source separation. It exploits information about the sources' fundamental frequencies (F0). The method integrates domain knowledge in the form of parametric source models into the DNN. Experimental evaluation shows that the proposed method outperforms F0-informed learning-free methods based on non-negative matrix factorization and a F0-informed supervised deep learning baseline. Moreover, the proposed method is extremely data-efficient. It makes powerful deep learning based source separation usable in domains where labeled training data is expensive or non-existent. (10.70675/1b805c5aze86az4508z8db4z9213ac926efb)
  DOI : 10.70675/1b805c5aze86az4508z8db4z9213ac926efb
• Using Decentralised Conflict-Abduction-Negation in Policy-making
  
  • Houze Etienne
  • Diaconescu Ada
  • Dessalles Jean-Louis
  , 2021.
• Multi-paradigm Modelling for Policy-driven Socio-technical Systems
  
  • Diaconescu Ada
  • Blouin Dominique
  • Ludvig Alice
  , 2021.
• Latency verification in execution traces of HW/SW partitioning model
  
  • Zoor Maysam
  , 2021. While many research works aim at defining new (formal) verification techniques to check for requirements in a model, understanding the root cause of a requirement violation is still an open issue for complex platforms built around software and hardware components. For instance, is the violation of a latency requirement due to unfavorable real-time scheduling, to contentions on buses, to the characteristics of functional algorithms or hardware components?This thesis introduces a Precise Latency ANalysis approach called PLAN. PLAN takes as input an instance of a HW/SW partitioning model, an execution trace, and a time constraint expressed in the following format: the latency between operator A and operator B should be less than a maximum latency value. First PLAN checks if the latency requirement is satisfied. If not, the main interest of PLAN is to provide the root cause of the non satisfaction by classifying execution transactions according to their impact on latency: obligatory transaction, transaction inducing a contention, transaction having no impact, etc.A first version of PLAN assumes an execution for which there is a unique execution of operator A and a unique execution of operator B. A second version of PLAN can compute, for each executed operator A, the corresponding operator B. For this, our approach relies on tainting techniques.The thesis formalizes the two versions of PLAN and illustrates them with toy examples. Then, we show how PLAN was integrated into a Model-Driven Framework (TTool). The two versions of PLAN are illustrated with two case studies taken from the H2020 AQUAS project. In particular, we show how tainting can efficiently handle the multiple and concurrent occurrences of the same operator. (10.70675/c3caad07z7061z4e9fzb24czd36ba279116d)
  DOI : 10.70675/c3caad07z7061z4e9fzb24czd36ba279116d
• Mixture weights optimisation for alpha-divergence variational inference
  
  • Daudel Kamélia
  • Douc Randal
  , 2021, 34, pp.4397--4408. This paper focuses on $\alpha$-divergence minimisation methods for Variational Inference. We consider the case where the posterior density is approximated by a mixture model and we investigate algorithms optimising the mixture weights of this mixture model by $\alpha$-divergence minimisation, without any information on the underlying distribution of its mixture components parameters. The Power Descent, defined for all $\alpha \neq 1$, is one such algorithm and we establish in our work the full proof of its convergence towards the optimal mixture weights when $\alpha < 1$. Since the $\alpha$-divergence recovers the widely-used exclusive Kullback-Leibler when $\alpha \to 1$, we then extend the Power Descent to the case $\alpha = 1$ and show that we obtain an Entropic Mirror Descent. This leads us to investigate the link between Power Descent and Entropic Mirror Descent: first-order approximations allow us to introduce the Rényi Descent, a novel algorithm for which we prove an $O(1/N)$ convergence rate. Lastly, we compare numerically the behavior of the unbiased Power Descent and of the biased Rényi Descent and we discuss the potential advantages of one algorithm over the other. (10.48550/arXiv.2106.05114)
  DOI : 10.48550/arXiv.2106.05114