Yassin El Hillali

Class imbalance complicates image-based driver behavior recognition when hazardous behaviors are underrepresented. This study evaluates how synthetic-data quantity affects classification performance across convolutional and transformer architectures. Using the six-class DMD-TFIW0 dataset, FLUX.1-dev with a pretrained realism-oriented low-rank adaptation module generates images of drinking, yawning, sleepy driving, and distracted driving. Generated images undergo visual screening and semantic assessment using Contrastive Language–Image Pretraining (CLIP). A ResNet50-based classifier, ResNet-18, and Swin Transformer are compared under three training conditions: oversampling with conventional augmentation, and conventional augmentation supplemented with moderate or extensive synthetic data. Synthetic images are used exclusively for training; validation and testing use fixed partitions containing only real images.Relative to baseline, extensive synthetic supplementation increases macro F1-score from 94.61% to 95.53% for ResNet-18 and from 94.82% to 95.96% for Swin Transformer, which achieves the highest accuracy of 96.72%. However, the ResNet50-based classifier performs best with moderate supplementation, reaching a macro F1-score of 94.74\%, compared with 93.34\% at baseline and 94.40\% with extensive supplementation. Moderate supplementation also lowers Swin Transformer's macro F1-score to 94.23\%, demonstrating a non-monotonic response to synthetic-data quantity. These findings show that targeted diffusion-based augmentation can improve recognition on real test images within this benchmark, but its benefits depend on both architecture and augmentation level. Synthetic-data quantity therefore requires architecture-specific evaluation.

The increasing use of cloud and collaborative computing creates a growing need to process sensitive data while keeping them encrypted. Existing encryption mechanisms either prevent direct computation or rely on general-purpose homomorphic constructions whose parameters and computational cost may exceed the requirements of structured applications with known operations and bounded computation depth. We address this problem by introducing a workload-driven computable symmetric encryption scheme whose parameters are derived directly from the structure of the data, the required operations, the numerical precision, and the targeted computation depth. The construction organizes text and numerical data into semantic coordinates and combines them through a hierarchical Chinese Remainder Theorem (CRT) representation. The main challenges are to preserve exact computation across the CRT hierarchy, prevent modular overflow, remove deterministic algebraic relations between stages, and introduce randomization without destroying the supported operations. To address them, we derive explicit capacity bounds, introduce inter-stage decorrelation, and develop two complementary protection profiles: a PRF-masked profile and a secret-lifting profile supporting bounded encrypted arithmetic. The resulting scheme enables targeted encrypted-text substitutions, matrix aggregation, Hadamard products, and collaborative or federated computations. We evaluate the construction through correctness, leakage-oriented security experiments, comparative benchmarks, and parameter-growth studies. The results confirm exact recovery within the prescribed bounds, effective suppression of the targeted structural leakages, efficient encrypted aggregation, and predictable cost growth with precision, data dimension, and computation depth.

Vehicular ad hoc networks (VANETs) require authentication that simultaneously guarantees message integrity, driver privacy, and traceability of malicious behaviour. Classical pseudonym public-key infrastructure (PKI) solutions rely on certificate pools, revocation lists, and frequent identity rotation, all of which become costly at scale. This paper proposes a zone-based anonymous authentication protocol founded on non-bearer challenge certificates. Each certificate is cryptographically linked to a driver–vehicle pair but requires an encrypted consistency proof to be used: road-side units (RSUs) store only 32-byte fingerprints rather than full certificate pools or revocation lists. For each zone, the on-board unit (OBU) selects a local secret, proves possession to a central authority (CA) via a Schnorr proof of knowledge, and retains the same secret—with domain-separated transcripts—for Diffie– Hellman establishment and vehicle-to-vehicle (V2V) relay signing; the CA issues a zone certificate without learning the secret. Authentication costs are shifted offline through pre-computation, reducing the online step at zone entry to a single hybrid encryption. Once authenticated, the RSU issues a pseudonymous local certificate enabling V2V relay through an append-only trace; key-agreement tokens remain ephemeral. Revocation operates at the pseudonymous layer; only the CA can lift the civil identity and pre-emptively cancel credentials for subsequent zones. The evaluation combines cryptographic micro-benchmarks, a radio-range calibration from an ITS toolchain (CARLA, SUMO, OMNeT++, Veins), and a Monte Carlo relay simulation at three densities. Geographic relay policies reduce the mean hop count from 3.76 to 3.18 (15%) and the mean authenticated packet from 2555 to 2243 bytes. Blockchain is not required for authentication.

The emergence of Vision-Language Models (VLMs) like Contrastive Language-Image Pretraining (CLIP) provides appealing solutions to various vision problems including Dynamic Facial Expression Recognition (DFER). However, most of the proposed approaches face major challenges, particularly related to inefficient full fine-tuning of the encoders and the complexity of the models. Moreover, some of the proposed methods seem to struggle with suboptimal performance due to (i) poor alignment between textual and visual representations, and (ii) ineffective temporal modeling. To address these challenges, we propose PE-CLIP, a parameter-efficient fine-tuning (PEFT) framework that elegantly adapts CLIP for dynamic facial expression recognition, requiring significantly reduced number of trainable parameters while maintaining high accuracy. At its core, to enhance efficiency and performance, PE-CLIP introduces two specialized adapters namely a Temporal Dynamic Adapter (TDA) and a Shared Adapter (ShA). The TDA is a GRU-based module with a dynamic scaling mechanism, capturing sequential dependencies while adaptively modulating the contribution of each temporal feature to emphasize the most informative ones while mitigating irrelevant variations. The ShA is a lightweight adapter refine representations within both textual and visual encoders, ensuring consistent feature processing while maintaining parameter efficiency. Additionally, we leverage Multi-modal Prompt Learning (MaPLe), which introduces learnable prompts to both visual and action unit-based textual description inputs, further improving the semantic alignment between modalities and enabling the efficient adaptation of CLIP for dynamic tasks. We evaluate our proposed PE-CLIP on two benchmark datasets, namely DFEW, FERV39K, and AFEW, achieving competitive performance compared to state-of-the-art methods while requiring fewer trainable parameters. By striking an optimal balance between parameter efficiency and performance, PE-CLIP sets a new benchmark in resource-efficient DFER. The source code of the proposed PE-CLIP will be publicly available at https://github.com/Ibtissam-SAADI/PE-CLIP .

Manufacturing enterprises implementing Industry 4.0 technologies face significant challenges in integrating heterogeneous maintenance data sources and deploying AI solutions effectively. While various AI methods exist for predictive maintenance, the fundamental challenge lies in creating a cohesive architecture that enables seamless data flow and AI deployment. This paper presents a standardized architecture framework with initial implementation steps at Toyota Motor Manufacturing France. The proposed architecture introduces a four-layer approach: (1) a unified data acquisition layer integrating IoT sensors, CMMS, and legacy systems through standardized interfaces (OPC UA/MQTT), (2) a data quality and standardization layer ensuring consistent formats and automated validation, (3) a modular AI deployment layer supporting anomaly detection (Wavelet-based analysis and Deep Learning) and remaining useful life prediction (LSTM networks), and (4) a maintenance workflow integration layer with bi-directional feedback. Key innovations include a unified maintenance data model, configurable data quality pipelines, and human-in-the-loop decision support. A conceptual validation suggests this architecture can improve integration efficiency and reduce equipment downtime. This research contributes to smart maintenance by providing a scalable architecture that balances interoperability, data quality, and practical deployment in brownfield environments.

Existing methods for driver facial expression recognition (DFER) are often computationally intensive, rendering them unsuitable for real-time applications. In this work, we introduce a novel transfer learning-based dual architecture, named ShuffViT-DFER, which elegantly combines computational efficiency and accuracy. This is achieved by harnessing the strengths of two lightweight and efficient models using convolutional neural network (CNN) and vision transformers (ViT). We efficiently fuse the extracted features to enhance the performance of the model in accurately recognizing the facial expressions of the driver. Our experimental results on two benchmarking and public datasets, KMU-FED and KDEF, highlight the validity of our proposed method for real-time application with superior performance when compared to state-of-the-art methods.