Fully Homomorphic Encryption (FHE) enables computations to be performed directly on encrypted data without requiring decryption, providing strong privacy guarantees. However, FHE remains computationally expensive, and writing efficient FHE programs is a complex, error-prone, and time-consuming task that demands significant cryptographic expertise. Programmers are often unaware of available optimizations, and applying them manually requires substantial effort. In this paper, we present CHEHAB, a compiler that automatically vectorizes scalar code, optimizes it, and generates highly efficient FHE programs. CHEHAB supports both structured and unstructured code and takes as input programs written in a domain-specific language embedded in C++. It relies on a Term Rewriting System (TRS) based on equality saturation to simplify and transform programs. CHEHAB targets two key challenges in FHE compilation: (1) automatic vectorization of scalar code, and (2) reduction of instruction execution latency and ciphertext noise growth. By leveraging equality saturation, CHEHAB explores a large optimization space to reduce instruction count and circuit depth while improving vector utilization. Experimental evaluation on a set of representative kernels shows that CHEHAB outperforms Coyote, a state-of-the-art vectorizing compiler for FHE. On average, CHEHAB generates code that is 7.38× faster at runtime, incurs 2.49× less accumulated noise, and achieves 251× faster compilation time. CHEHAB is released as an open-source compiler to support reproducibility and further research in FHE compilation.

Heterogeneous AI accelerators that combine high-precision digital cores with energy-efficient analog in-memory computing (AIMC) units offer a promising path to overcome the energy and scalability limits of deep learning. A key challenge, however, is to determine which neural network layers can be executed on noisy analog units without compromising accuracy. Existing mapping strategies rely on ad-hoc heuristics and lack principled noise-sensitivity estimation. We propose HILAL (Hessian-Informed Layer Allocation), a framework that systematically quantifies layer robustness to analog noise using two complementary metrics: Hessian-based Noise Impact and Spectral Concentration Ratio. Layers are partitioned into robust and sensitive groups via clustering, enabling threshold-free mapping to analog or digital units. To further mitigate accuracy loss, we gradually offload layers to AIMC while retraining with noise-injection. Experiments on convolutional networks and transformers across CIFAR-10/100, ImageNet and SQuAD show that HILAL is on average 3.09x faster in search and mapping runtime than SOTA methods while achieving less accuracy degradation and maximizing analog utilization.

While polyhedral compilers have shown success in implementing advanced code transformations, they still face challenges in selecting the ones that lead to the most profitable speedups. This has motivated the use of machine learning based cost models to guide the search for polyhedral optimizations. State-of-the-art polyhedral compilers have demonstrated a viable proof-of-concept of such an approach. While promising, this approach still faces significant limitations. Existing polyhedral compilers using deep learning cost models typically support only a small subset of affine transformations, limiting their ability to explore complex code transformations. Furthermore, their applicability does not scale beyond simple programs, thus excluding many program classes from their scope, such as those with non-rectangular iteration domains or multiple loop nests. These limitations significantly impact the generality of such compilers and autoschedulers, raising questions about the overall approach. In this paper, we introduce LOOPER, the first polyhedral autoscheduler that uses a deep learning based cost model and covers a large space of affine transformations and programs. LOOPER allows the optimization of an extensive set of programs while being effective at applying complex sequences of polyhedral transformations. We implement and evaluate LOOPER and show that it achieves competitive speedups over the state-of-the-art. On the PolyBench benchmarks, LOOPER achieves a geometric mean speedup of $\mathbf{1 . 8 4} \mathbf{x}$ over the Tiramisu autoscheduler and $\mathbf{1 . 4 2} \mathbf{x}$ over Pluto, two state-of-the-art polyhedral autoschedulers.

We address the Unrelated Parallel Machine Scheduling Problem (UPMSP), where machines are subject to unavailability periods due to maintenance interventions. Two conflicting objectives are considered: the first focuses on production and maintenance tardiness, while the second aims to minimize energy consumption. This problem was previously studied in [15], where it was formulated as a Mixed-Integer Linear Programming (MILP) model and solved using an enhanced Multi-Objective Simulated Annealing algorithm based on a population of solutions (POP-MOSA).In this work, we aim to improve the performance of POP-MOSA by integrating reinforcement learning techniques, specifically the Q-learning algorithm. The resulting approach, referred to as Q-MOSA, leverages Q-learning to adaptively select the most appropriate local search strategy at each iteration.Q-MOSA introduces several novel contributions. First, at the end of the learning phase, the Q-table is generated using a compact design that reduces its size while preserving decision quality. Second, given the conflicting nature of minimizing both tardiness and energy consumption, a more flexible learning mechanism is required. This motivates the adoption of a multi-reward reinforcement learning strategy, which offers a more effective alternative to traditional single-reward schemes.Q-MOSA is evaluated using small-scale benchmark instances, and its performance is compared against the exact Pareto front obtained in [15].

Homomorphic encryption (HE) enables secure computation on encrypted data, making privacy-preserving machine learning (ML) viable for sensitive applications. Non-linear activation functions (AFs) such as ReLU, Sigmoid, and Tanh remain costly, as traditional polynomial approximations incur deep circuits and long runtimes. We propose a GPU-accelerated method that approximates these activations using fast piecewise-linear functions, formulated as a sign-decision problem to simplify computation and reduce circuit complexity. Compared to existing methods, our technique achieves a lower mean squared error (MSE) to the true activation functions, reducing MSE by approximately $55 \%$ with an error below $10^{-2}$. Experiments demonstrate speedups of up to $128 \times$ on MNIST and $840 \times$ on CIFAR-10, with accuracy comparable to state-of-the-art methods. Across various neural architectures and datasets, our pipeline significantly reduces homomorphic evaluation time and circuit depth compared to polynomial approximations of similar accuracy.

Providing high-quality predictions is no longer the sole goal for neural networks. As we live in an increasingly interconnected world, these models need to match the constraints of resource-limited devices powering the Internet of Things (IoT) and embedded systems. Moreover, in the era of climate change, reducing the carbon footprint of neural networks is a critical step for green artificial intelligence, which is no longer an aspiration but a major need. Enhancing the energy efficiency of neural networks, in both training and inference phases, became a predominant research topic in the field. Training optimization has grown in interest recently but remains challenging, as it involves changes in the learning procedure that can impact the prediction quality significantly. This article presents a study on the most popular techniques aiming to reduce the energy consumption of neural networks' training. We first propose a classification of the methods before discussing and comparing the different categories. In addition, we outline some energy measurement techniques. We discuss the limitations identified during our study as well as some interesting directions, such as neuromorphic and reservoir computing (RC).

The objective of this study is to address a scheduling issue arising in production workshops, specifically dealing with unrelated parallel machine scheduling integrating flexible and periodic maintenance interventions. Machines, while operational, consume varying amounts of energy based on their state: idle states consume less energy than active ones. Additionally, energy consumption is influenced by both the production job and the specific machine.Our goal is to minimize two functions: the first pertains to reducing production and maintenance earliness/tardiness, while the second focuses on minimizing energy consumption. To address this problem, a Mixed Integer Linear Program (MILP) is proposed. The ϵ—constrained method is employed for computing the Pareto front, and further adaptation involves applying the Multi-Objective Simulated Annealing algorithm (MOSA) to handle instances of substantial size.The proposed MILP model demonstrates the ability to accurately determine the Pareto front for up to 30 production jobs and 2 machines on literature benchmarks within a timeframe of less than 1 hour. Moreover, the computed metrics proves the effectiveness of the proposed MOSA.

The introduction of 6G technology into the Internet of Vehicles (IoV) promises to revolutionize connectivity with ultra-high data rates and seamless network coverage. However, this technological leap also brings significant challenges, particularly for the dynamic and diverse IoV landscape, which must meet the rigorous reliability and security requirements of 6G networks. Furthermore, integrating 6G will likely increase the IoV’s susceptibility to a spectrum of emerging cyber threats. Therefore, it is crucial for security mechanisms to dynamically adapt and learn new attack patterns, keeping pace with the rapid evolution and diversification of these threats - a capability currently lacking in existing systems. This paper presents a novel intrusion detection system leveraging the paradigm of life-long (or continual) learning. Our methodology combines class-incremental learning with federated learning, an approach ideally suited to the distributed nature of the IoV. This strategy effectively harnesses the collective intelligence of Connected and Automated Vehicles (CAVs) and edge computing capabilities to train the detection system. To the best of our knowledge, this study is the first to synergize class-incremental learning with federated learning specifically for cyber attack detection. Through comprehensive experiments on a recent network traffic dataset, our system has exhibited a robust adaptability in learning new cyber attack patterns, while effectively retaining knowledge of previously encountered ones. Additionally, it has proven to maintain high accuracy and a low false positive rate.

Monitoring pollution plumes is a key issue, given the harmful effects they cause. The dynamic of these plumes, which may be important due to meteorological conditions, makes their study difficult. Real-time monitoring in order to obtain an accurate mapping of the pollution dispersion is helpful and valuable to mitigate risks. In this work, we consider a fleet of cooperative drones carrying pollution sensors and operating in order to assess a pollution plume. The latter is assumed to follow a Gaussian process (GP) with varying parameters. For this use case, we propose an efficient approach to characterize spatially and temporarily the plume while optimizing the path planning of drones. In our approach, drones are guided by a deep reinforcement learning (DRL) model called the categorical deep$Q$-network (Categorical DQN) to maximize the plume coverage while considering budget constraints. Specifically, we develop a scalable independent$Q$-learning (IQL) scheme that shares team rewards based on each drone’s deployment relevance and, therefore, ensures cooperation. We evaluate the performance of the plume parameter estimation as well as the maps generated by the GP regression. By testing our framework on several plume scenarios, we show that it offers good results in terms of both estimation quality and runtime efficiency.

As automated decision-making systems proliferate, accountability becomes crucial. Developers must ensure adherence to regulations and fairness. Explainable AI offers a remedy by crafting algorithms that provide precise outcomes and understandable explanations. This paper focuses on food recommender system interpretability for better health. Integrating explainable AI empowers users to make informed dietary decisions. The proposed framework generates natural language explanations for recommendations using the prompting technique, demonstrating superior performance and broad applicability across domains.

In this paper, we address the energy efficiency of neural networks training through data selection techniques. We first study the impact of a random data selection approach that renews the selected examples periodically during training. We find that random selection should be considered as a serious option as it allows high energy gains with small accuracy losses. Unexpectedly, it even outperforms a more elaborate approach in some cases.Our study of the random approach conducted us to observe that low selectivity ratios allow important energy savings, but also cause a significant accuracy decrease. To mitigate the effect of such ratios on the prediction quality, we propose to use a dynamic selectivity ratio with a decreasing schedule, that can be integrated to any selection approach. Our first results show that using such a schedule provides around 60% energy gains on the CIFAR-10 dataset with less than 1% accuracy decrease. It also improves the convergence when compared to a fixed ratio.

In this paper, we analyze the results of two hyper-heuristics HHGA and HHabs that generate per-instances genetic algorithms for the permutation flow shop problem. They are competitive with literature approaches for most of instances of the benchmark of Taillard. Nevertheless, they are not effective enough for some difficult instances. For this purpose, we propose a workflow to analyse GAs configurations and their results in order to detect which components influence the most on generated GAs quality in order to enhance the quality of these hyper-heuristics.

Hyper-heuristics can be applied to solve complex optimization problems. However, they need a substantial number of fitness function evaluations to discover a good approximation to the global optimum, especially for large-scale problems. Recently, surrogate-assisted algorithms have drawn increasing attention, and have shown their potential to deal with expensive complex optimization problems. This paper aims to use surrogates to approximate HHGA's (Ahmed Bacha et al., 2019) fitness functions, an efficient hyper-heuristic for solving the permutation flowshop problem, one of the most important scheduling types in modern industries. The objective is to approximate, in an online approach, the fitness function, reducing considerably the execution time of HHGA while maintaining its quality. The proposed online surrogate model is mainly designed to capture the details of the fitness function to enhance the accuracy estimation. The experimental results on Taillard's widely used benchmark problems show that the proposed fitness approximation-assisted HHGA is able to achieve competitive performance on a limited computational budget.

In this paper, we propose a hybrid Artificial Bee Colony algorithm with Simulated Annealing (ABC-SA) to address the community detection problem. SA enhances the exploitation by searching the most promising regions located by ABC algorithm. Besides, in order to accommodate the characteristics of social networks, we use locus-based adjacency encoding scheme, in which communities are identified as a graph connected components and Pearson's correlation as structural information to guide the solutions' construction. Results obtained on synthetic and real-word networks show that the proposed algorithm can discover communities more successfully in comparison with traditional ABC algorithm and other state-of-the-art algorithms.

Recently, palmprint recognition in forensic domain has gained considerable attention since 30 % of evidence left in crime scene originate from palms. Like most of recognition systems, palmprint one is composed of three steps: preprocessing, features extraction and representation and finally features matching. Minutiae are the most reliable and discriminating features used in these systems. Minutiae matching is then very critical. Quantifying the similarity between two sets of extracted minutiae and assigning a score is particularly important in this step. In this paper, we designed similarity scores for a minutiae based recognition system using a minimum of extracted information. Our proposed scores are based on the score of [1], used in point pattern matching. They are tested and compared on the database used in [2]. The best one is tested and compared to the one presented in the same work [2]. Obtained results are very interesting.

This paper deals with induction machines bearing failures detection and diagnosis using vibration and temperature signals. The failure detection is managed by a clustering graphical representation creating transition classes. Motivated by the computational complexity of the problem, a Variable Neighborhood Search (VNS) metaheuristic is developed including well-designed local search algorithms for data clustering to the system diagnosis. Computational experiments carried out on the PRONOSTIA experimental platform data show that the proposed algorithm seems to be efficient and effective.

In this paper, we propose HHGA, a new hyper-heuristic for the Permutation Flowshop Problem with makespan minimization. It consists of a high-level genetic algorithm which goal is to tailor a dedicated and effective genetic algorithm for each instance. The goal of the proposed hyper-heuristic is to find among a set of genetic operators, a configuration that is suitable for solving a PFSP problem instance and compare the tailored GAs to investigate the existence of any patterns. Our experiments on well-known Taillard’s benchmark showed the performance of tailored GAs. HHGA was able to build genetic algorithms that reached optimal solutions in 50 instances out of 120 and was at least as good as state-of-the-art approaches. Through results, we show also the performance of some operators compared to others.

One of the most interesting issues in the field of social network analysis is community evolution prediction in dynamic social networks. To start with, the dynamic network is split into a series of timeframes, each one containing interactions aggregated over a time period such as a month, a day or an hour. Splitting the network into timeframes is of crucial importance to capture the right communities' temporal evolution before predicting their future. Our paper investigates the problem of choosing the appropriate scale for network splitting which would improve the prediction. The experiments we conducted on Facebook and Higgs Twitter datasets offer strong empirical evidence of the usefulness of considering the appropriate network splitting as a first step in predicting community evolution in dynamic social networks.

This paper deals with induction machines bearing failures detection and diagnosis using vibration and temperature signals. It proposes the use of a new Classical Mechanics-inspired Optimization (CMO) metaheuristic for data clustering. To ensure failure detection, transitions from a state to another is analyzed in order to form a transitional model between system states generated by the clustering. The performances of the proposed new metaheuristic are evaluated on the PRONOSTIA experimental platform data.

The work described in this paper is about building a general model capable of simulating human behaviour and emotions using virtual characters. To make the simulation realistic enough, virtual characters need to express emotions according to the environment and deal with those emotions in a parallel way where an emotional experience can be triggered at the same time as another emotional response. The virtual character will have perceptions, feel and express emotions and respond to different situations. To make the simulation realistic, we used a method allowing the virtual characters to execute tasks, perceive events and display emotions in a parallel way. To do that, we used the multiple resources model [1] to control the attention and to predict when two or more actions can be executed at the same time. The used emotional model is based on Scherer's theory [2]. However, in this paper we focus on the control of the attention as a part of the emotional process.

In this paper, we introduce Adaptive Data-Communication Trust Mechanism for clustered wireless sensor networks (ADCT) to effectively deal with compromise or malicious nodes. Unlike prior works, we propose an adaptive function to evaluate the direct trust between nodes according to the need of the application in terms of trust severity. We also consider data trust to cope with untrustworthy nodes during the data collection despite their communication capabilities. Finally, we introduce a new method to discard a malicious recommendation when building feedback at the base station or cluster-head level. Theoretical analysis shows that our trust mechanism allows a good trade-off between memory requirement and trust cooperation.

Palmprint is one of the modalities that offer high recognition accuracy. The recognition process depends on an optimized ROI (Region of Interest) extraction. This extraction is affected by several factors including the device used and the acquisition conditions. The acquisition mode can alter some image properties like rotation, translation and scale. Some devices are designed to maintain hand in a fixed position and delimit a subspace of the hand. On the other hand, contactless devices offer more convenience and flexibility but lead to altered images. ROI extraction methods must consider the acquisition device (with contact or contactless). In this paper, we propose a ROI extraction method that addresses this issue. We test our method on two databases PolyU and CASIA which illustrate the impact of using contactless device unlike the PolyU device. Then, we test performances of the palmprint biometric system. We use a Fisher Linear Discriminant projection (FLD) to extract features from ROI transformed into the frequency domain. Our proposed method can significantly cover a great portion of the palm in the two databases. Performances obtained with the proposed palmprint system are promising.

Detection of moving objects in a video sequence is a growing field, used in various domains. There are, in the literature, several approaches to detect such movements. In this paper, we will focus on one of these approaches namely block matching. We propose to speed the performances of the block matching algorithm using the Bees' Algorithm which is known for its efficient exploration of search space as it uses two interesting strategies intensification and diversification. The proposed algorithm, BAforBM, is evaluated and compared to two existing ones: genetic bloc matching algorithm and a block matching algorithm based on PSO (Particule Swarm Optimisation).

Tree topologies are widely used in WSN in order to route convergecast traffic to the sink. We consider in this paper the Shortest Path routing Tree (SPT) problem in WSN under different metrics; we show that the basic SPT based strategies are unsuitable for the many-to-one WSN when considering some metrics to compute link costs. Indeed, existing SPT approaches aim to construct a tree rooted at the sink such that the cost of the path from any node to the sink is minimal, while the cost of a given path is computed as summation of the costs of links that compose this path. However, in many-to-one WSN, links which are close to the sink are more critical than other links when using some metrics. We propose in this paper a new weighted path cost function, and we show that our cost function is more suitable for WSN. Based on this cost function, we propose a simple and efficient weighted shortest path tree construction which does not introduce new overheads. We consider, then, the particular case of energy-aware routing in WSN when we apply our new solution in order to construct more suitable energy-aware SPT. We conduct extensive simulations which show that our approach allows to enhance the network lifetime up to 17% compared to the basic one.