Additive margin in contrastive self-supervised frameworks to learn discriminative speaker representations
In The speaker and language recognition workshop (odyssey 2024)
Abstract
Self-Supervised Learning (SSL) frameworks became the standard for learning robust class representations by benefiting from large unlabeled datasets. For Speaker Verification (SV), most SSL systems rely on contrastive-based loss functions. We explore different ways to improve the performance of these techniques by revisiting the NT-Xent contrastive loss. Our main contribution is the definition of the NT-Xent-AM loss and the study of the importance of Additive Margin (AM) in SimCLR and MoCo SSL methods to further separate positive from negative pairs. Despite class collisions, we show that AM enhances the compactness of same-speaker embeddings and reduces the number of false negatives and false positives on SV. Additionally, we demonstrate the effectiveness of the symmetric contrastive loss, which provides more supervision for the SSL task. Implementing these two modifications to SimCLR improves performance and results in 7.85% EER on VoxCeleb1-O, outperforming other equivalent methods.
Apprentissage interprétable de la criminalité en france (2012-2021)
In Actes de l’atelier gestion et analyse des données spatiales et temporelles
Abstract
L’activité criminelle en France a connu une évolution significative au cours des deux dernières décennies, marquée par la recrudescence des actes de malveillance, notamment liés aux mouvements sociaux et syndicaux, aux émeutes, ainsi qu’au terrorisme. Dans ce contexte difficile, l’utilisation de techniques issues de l’intelligence artificielle pourrait offrir de nombreuses perspectives pour renforcer la sûreté publique et privée en France. Un exemple de cette approche est l’analyse spatio-temporelle des données de criminalité, déjà couronnée de succès au Brésil (Da Silva et al., 2020), au Proche-Orient (Tolan et al., 2015), et dans d’autres pays. Dans le cadre de ce travail, nous explorons la possibilité d’appliquer cette approche au contexte français.
Enhanced neonatal screening for sickle cell disease: Human-guided deep learning with CNN on isoelectric focusing images
In Journal of Infrastructure, Policy and Development
Abstract
Accurate detection of abnormal hemoglobin variations is paramount for early diagnosis of sickle cell disease (SCD) in newborns. Traditional methods using isoelectric focusing (IEF) with agarose gels are technician-dependent and face limitations like inconsistent image quality and interpretation challenges. This study proposes a groundbreaking solution using deep learning (DL) and artificial intelligence (AI) while ensuring human guidance throughout the process. The system analyzes IEF gel images with convolutional neural networks (CNNs), achieving over 98% accuracy in identifying various SCD profiles, far surpassing the limitations of traditional methods. Furthermore, the system addresses ambiguities by incorporating an “Unconfirmed” category for unclear cases and assigns probability values to each classification, empowering clinicians with crucial information for informed decisions. This AI-powered tool, named SCScreen, seamlessly integrates machine learning with medical expertise, offering a robust, efficient, and accurate solution for SCD screening. Notably, SCScreen tackles the previously challenging diagnosis of major sickle cell syndromes (SDM) in newborns. This research has the potential to revolutionize SCD management. By strengthening screening platforms and potentially reducing costs, SCScreen paves the way for improved healthcare outcomes for newborns with SCD, potentially saving lives and improving the quality of life for affected individuals.
Exploring WavLM back-ends for speech spoofing and deepfake detection
In ASVSpoof interspeech 2024
Abstract
This paper describes our submitted systems to the ASVspoof 5 Challenge Track 1: Speech Deepfake Detection - Open Condition, which consists of a stand-alone speech deepfake (bonafide vs spoof) detection task. Recently, large-scale self-supervised models become a standard in Automatic Speech Recognition (ASR) and other speech processing tasks. Thus, we leverage a pre-trained WavLM as a front-end model and pool its representations with different back-end techniques. The complete framework is fine-tuned using only the trained dataset of the challenge, similar to the close condition. Besides, we adopt data-augmentation by adding noise and reverberation using MUSAN noise and RIR datasets. We also experiment with codec augmentations to increase the performance of our method. Ultimately, we use the Bosaris toolkit for score calibration and system fusion to get better Cllr scores. Our fused system achieves $0.0937$ minDCF, $3.42%$ EER, $0.1927$ Cllr, and $0.1375$ actDCF.
Towards supervised performance on speaker verification with self-supervised learning by leveraging large-scale ASR models
In Interspeech 2024
Abstract
Recent advancements in Self-Supervised Learning (SSL) have shown promising results in Speaker Verification (SV). However, narrowing the performance gap with supervised systems remains an ongoing challenge. Several studies have observed that speech representations from large-scale ASR models contain valuable speaker information. This work explores the limitations of fine-tuning these models for SV using an SSL contrastive objective in an end-to-end approach. Then, we propose a framework to learn speaker representations in an SSL context by fine-tuning a pre-trained WavLM with a supervised loss using pseudo-labels. Initial pseudo-labels are derived from an SSL DINO-based model and are iteratively refined by clustering the model embeddings. Our method achieves 0.99% EER on VoxCeleb1-O, establishing the new state-of-the-art on self-supervised SV. As this performance is close to our supervised baseline of 0.94% EER, this contribution is a step towards supervised performance on SV with SSL.
A CP-based automatic tool for instantiating truncated differential characteristics
In Progress in cryptology – INDOCRYPT 2023
Abstract
An important criteria to assert the security of a cryptographic primitive is its resistance against differential cryptanalysis. For word-oriented primitives, a common technique to determine the number of rounds required to ensure the immunity against differential distinguishers is to consider truncated differential characteristics and to count the number of active S-boxes. Doing so allows to provide an upper bound on the probability of the best differential characteristic with a reduced computational cost. However, in order to design very efficient primitives, it might be needed to evaluate the probability more accurately. This is usually done in a second step, during which one tries to instantiate truncated differential characteristics with actual values and computes its corresponding probability. This step is usually done with ad-hoc algorithms and CP or MILP models for generic solvers. In this paper, we present a generic approach for automatically generating these models to handle all word-oriented ciphers. Furthermore the running times to solve these models are very competitive with all the previous dedicated approaches.