• Developed Monitoring tool collectors for fine-grained edge workload monitoring • Lightweight pipeline for intent-based assurance on resource-constrained devices • Real-time detection of application transitions using an autoencoder model • Fast and accurate Application Type Identification via Random Forest classifier • Public AIMED-2025 dataset with 9 workloads on RPi to support the research community Intent-Based Networking (IBN) enables agile and policy-driven network management by translating high-level intents into concrete configurations and continuously validating their compliance. A critical limitation in current Intent-Based Network Assurance (IBNA) systems is the lack of real-time application-level awareness, particularly in dynamic edge environments where AI workloads frequently change. In this work, we address this limitation by introducing a lightweight, monitoring-driven pipeline that enables the detection of application transitions and identification of newly active application types on edge devices. In collaboration with Netdata engineers, we develop multimetric data collectors using Netdata, an open-source platform for real-time system and application monitoring. These collectors capture application-agnostic system metrics with minimal overhead, forming the foundation for real-time alerting and dynamic network adaptation. Our proposed pipeline transforms raw monitoring data into fixed-length vectorized multivariate time series. An undercomplete autoencoder is then used to detect changes in system behavior indicative of application transitions, followed by a Random Forest classifier that labels the newly active application based on its resource usage profile. To support reproducibility, we construct and publicly release the AIMED-2025 dataset, which includes monitoring data from seven MediaPipe-based edge AI applications and two idle states, all executed on a Raspberry Pi. Experimental evaluation demonstrates that our method achieves 100% accuracy in both Application Transition Detection and Application Type Identification using only a three-second observation window. Furthermore, the system exhibits sub-second training times and millisecond-scale inference latency, making it suitable for real-time deployment on resource-constrained edge devices. Once an application change is detected and identified, the IBNA system can automatically alert network administrators and trigger dynamic reconfiguration of network resources to meet the specific performance, security, and connectivity requirements of the active application. By integrating application-level awareness into IBNA, this work advances the state of the art in intent-driven network management and enables more adaptive, efficient, and reliable operation of edge AI systems.

Domain Generalization (DG) research has gained considerable traction as of late, since the ability to generalize to unseen data distributions is a requirement that eludes even state-of-the-art training algorithms. In this paper we observe that the initial iterations of model training play a key role in domain generalization effectiveness, since the loss landscape may be significantly different across the training and test distributions, contrary to the case of i.i.d. data. Conflicts between gradients of the loss components of each domain lead the optimization procedure to undesirable local minima that do not capture the domain-invariant features of the target classes. We propose alleviating domain conflicts in model optimization, by iteratively annealing the parameters of a model in the early stages of training and searching for points where gradients align between domains. By discovering a set of parameter values where gradients are updated towards the same direction for each data distribution present in the training set, the proposed Gradient-Guided Annealing (GGA) algorithm encourages models to seek out minima that exhibit improved robustness against domain shifts. The efficacy of GGA is evaluated on five widely accepted and challenging image classification domain generalization benchmarks, where its use alone is able to establish highly competitive or even state-of-the-art performance. Moreover, when combined with previously proposed domain-generalization algorithms it is able to consistently improve their effectiveness by significant margins1.

Mitigating biases in computer vision models is an essential step towards the trustworthiness of artificial intelligence models. Existing bias mitigation methods focus on a small set of predefined biases, limiting their applicability in visual datasets where multiple, possibly unknown biases exist. To address this limitation, we introduce MAVias, an open-set bias mitigation approach leveraging foundation models to discover spurious associations between visual attributes and target classes. MAVias first captures a wide variety of visual features in natural language via a foundation image tagging model, and then leverages a large language model to select those visual features defining the target class, resulting in a set of language-coded potential visual biases. We then translate this set of potential biases into vision-language embeddings and introduce an in-processing bias mitigation approach to prevent the model from encoding information related to them. Our experiments on diverse datasets, including CelebA, Waterbirds, ImageNet, and UrbanCars, show that MAVias effectively detects and mitigates a wide range of biases in visual recognition tasks outperforming current state-of-the-art.

Bias in computer vision systems can perpetuate or even amplify discrimination against certain populations. Considering that bias is often introduced by biased visual datasets, many recent research efforts focus on training fair models using such data. However, most of them heavily rely on the availability of protected attribute labels in the dataset, which limits their applicability, while label-unaware approaches, i.e., approaches operating without such labels, exhibit considerably lower performance. To overcome these limitations, this work introduces FLAC, a methodology that minimizes mutual information between the features extracted by the model and a protected attribute, without the use of attribute labels. To do that, FLAC proposes a sampling strategy that highlights underrepresented samples in the dataset, and casts the problem of learning fair representations as a probability matching problem that leverages representations extracted by a bias-capturing classifier. It is theoretically shown that FLAC can indeed lead to fair representations, that are independent of the protected attributes. FLAC surpasses the current state-of-the-art on Biased-MNIST, CelebA, and UTKFace, by 29.1%, 18.1%, and 21.9%, respectively. Additionally, FLAC exhibits 2.2% increased accuracy on ImageNet-A and up to 4.2% increased accuracy on Corrupted-Cifar10. Finally, in most experiments, FLAC even outperforms the bias label-aware state-of-the-art methods.

Advances in IoT technologies combined with new algorithms have enabled the collection and processing of high-rate multi-source data streams that quantify human behavior in a fine-grained level and can lead to deeper insights on individual behaviors as well as on the interplay between behaviors and the environment. In this paper, we present an integrated system that collects and extracts multiple behavioral and environmental indicators, aiming at improving public health policies for tackling obesity. Data collection takes place using passive methods based on smartphone and smartwatch applications that require minimal interaction with the user. Our goal is to present a detailed account of the design principles, the implementation processes, and the evaluation of integrated algorithms, especially given the challenges we faced, in particular (a) integrating multiple technologies, algorithms, and components under a single, unified system, and (b) large scale (Big Data) requirements. We also present evaluation results of the algorithms on datasets (public for most cases) such as an absolute error of 8-9 steps when counting steps, 0.86 F1-score for detecting visited locations, and an error of less than 12 mins for gross sleep time. Finally, we also briefly present studies that have been materialized using our system, thus demonstrating its potential value to public authorities and individual researchers.

EXplainable Artificial Intelligence (XAI) approaches are widely applied for identifying fairness issues in Artificial Intelligence (AI) systems. However, in the context of facial analysis, existing XAI approaches, such as pixel attribution methods, offer explanations for individual images, posing challenges in assessing the overall behavior of a model, which would require labor-intensive manual inspection of a very large number of instances and leaving to the human the task of drawing a general impression of the model behavior from the individual outputs. Addressing this limitation, we introduce FaceX, the first method that provides a comprehensive understanding of face attribute classifiers through summary model explanations. Specifically, FaceX leverages the presence of distinct regions across all facial images to compute a region-level aggregation of model activations, allowing for the visualization of the model's region attribution across 19 predefined regions of interest in facial images, such as hair, ears, or skin. Beyond spatial explanations, FaceX enhances interpretability by visualizing specific image patches with the highest impact on the model's decisions for each facial region within a test benchmark. Through extensive evaluation in various experimental setups, including scenarios with or without intentional biases and mitigation efforts on four benchmarks, namely CelebA, FairFace, CelebAMask-HQ, and Racial Faces in the Wild, FaceX demonstrates high effectiveness in identifying the models' biases.

Abstract In the realm of online continual learning, models are expected to adapt to an ever-changing environment. One of the most persistent hurdles in this adaptation is the mitigation of a phenomenon called "Catastrophic Forgetting" (CF). This critical condition occurs when models trained on non-identically distributed data lose performance in previously learned tasks. Rehearsal methods, leveraging the ability to replay older samples, aim to address this challenge by incorporating a buffer of past training samples. However, the absence of known task boundaries complicates the adaptation of current CF mitigation methods. This paper proposes a method attuned to data stream characteristics and online model performance in a resource-constrained environment. The number of training iterations and learning rate emerges as crucial hyperparameters, impacting the efficacy and efficiency of online continual learning. Up to this point, we propose a combination of Experience Replay methodologies, a Drift Detector, and various training convergence policies, specially tailored for scenarios with unknown task boundaries. Experimental results demonstrate the effectiveness of our approach, maintaining or enhancing performance compared to baseline methods, while significantly improving computational efficiency.

Accumulated Local Effects (ALE) is a widely-used explainability method for isolating the average effect of a feature on the output, because it handles cases with correlated features well. However, it has two limitations. First, it does not quantify the deviation of instance-level (local) effects from the average (global) effect, known as heterogeneity. Second, for estimating the average effect, it partitions the feature domain into user-defined, fixed-sized bins, where different bin sizes may lead to inconsistent ALE estimations. To address these limitations, we propose Robust and Heterogeneity-aware ALE (RHALE). RHALE quantifies the heterogeneity by considering the standard deviation of the local effects and automatically determines an optimal variable-size bin-splitting. In this paper, we prove that to achieve an unbiased approximation of the standard deviation of local effects within each bin, bin splitting must follow a set of sufficient conditions. Based on these conditions, we propose an algorithm that automatically determines the optimal partitioning, balancing the estimation bias and variance. Through evaluations on synthetic and real datasets, we demonstrate the superiority of RHALE compared to other methods, including the advantages of automatic bin splitting, especially in cases with correlated features.

In the past decade, deep convolutional neural networks have achieved significant success in image classification and ranking, finding therefore numerous applications in multimedia content retrieval. Still, these models suffer from performance degradation when neural networks are tested on out-of-distribution scenarios or on data originating from previously unseen data Domains. In the present work, we focus on this problem of Domain Generalization and propose an alternative neural network architecture for robust, out-of-distribution image classification. We attempt to produce a model that focuses on the causal features of the depicted class for robust image classification in the Domain Generalization setting. To achieve this, we propose attending to multiple-levels of information throughout a Convolutional Neural Network and leveraging the most important attributes of an image, by employing trainable attention mechanisms. To validate our method we evaluate our model on four widely accepted Domain Generalization benchmarks, where our model is able to surpass previously reported baselines in three out of four datasets and achieve the second best score in the fourth one.

Treatment effect estimation is of high-importance for both researchers and practitioners across many scientific and industrial domains. The abundance of observational data makes them increasingly used by researchers for the estimation of causal effects. However, these data suffer from several weaknesses, leading to inaccurate causal effect estimations, if not handled properly. Therefore, several machine learning techniques have been proposed, most of them focusing on leveraging the predictive power of neural network models to attain more precise estimation of causal effects. In this work, we propose a new methodology, named Nearest Neighboring Information for Causal Inference (NNCI), for integrating valuable nearest neighboring information on neural network-based models for estimating treatment effects. The proposed NNCI methodology is applied to some of the most well established neural network-based models for treatment effect estimation with the use of observational data. Numerical experiments and analysis provide empirical and statistical evidence that the integration of NNCI with state-of-the-art neural network models leads to considerably improved treatment effect estimations on a variety of well-known challenging benchmarks.

Oblivious inference is the task of outsourcing a ML model, like neural-networks, without disclosing critical and sensitive information, like the model's parameters. One of the most prominent solutions for secure oblivious inference is based on a powerful cryptographic tools, like Homomorphic Encryption (HE) and/or multi-party computation (MPC). Even though the implementation of oblivious inference systems schemes has impressively improved the last decade, there are still significant limitations on the ML models that they can practically implement. Especially when both the ML model and the input data's confidentiality must be protected. In this paper, we introduce the notion of partially oblivious inference. We empirically show that for neural network models, like CNNs, some information leakage can be acceptable. We therefore propose a novel trade-off between security and efficiency. In our research, we investigate the impact on security and inference runtime performance from the CNN model's weights partial leakage. We experimentally demonstrate that in a CIFAR-10 network we can leak up to $80\%$ of the model's weights with practically no security impact, while the necessary HE-mutliplications are performed four times faster.

The increased worldwide prevalence of obesity has sparked the interest of the scientific community towards tools that objectively and automatically monitor eating behavior. Despite the study of obesity being in the spotlight, such tools can also be used to study eating disorders (e.g. anorexia nervosa) or provide a personalized monitoring platform for patients or athletes. This paper presents a complete framework towards the automated i) modeling of in-meal eating behavior and ii) temporal localization of meals, from raw inertial data collected in-the-wild using commercially available smartwatches. Initially, we present an end-to-end Neural Network which detects food intake events (i.e. bites). The proposed network uses both convolutional and recurrent layers that are trained simultaneously. Subsequently, we show how the distribution of the detected bites throughout the day can be used to estimate the start and end points of meals, using signal processing algorithms. We perform extensive evaluation on each framework part individually. Leave-one-subject-out (LOSO) evaluation shows that our bite detection approach outperforms four state-of-the-art algorithms towards the detection of bites during the course of a meal (0.923 F1 score). Furthermore, LOSO and held-out set experiments regarding the estimation of meal start/end points reveal that the proposed approach outperforms a relevant approach found in the literature (Jaccard Index of 0.820 and 0.821 for the LOSO and held-out experiments, respectively). Experiments are performed using our publicly available FIC and the newly introduced FreeFIC datasets.

Overweight and obesity are both associated with in-meal eating parameters such as eating speed. Recently, the plethora of available wearable devices in the market ignited the interest of both the scientific community and the industry toward unobtrusive solutions for eating behavior monitoring. In this paper, we present an algorithm for automatically detecting the in-meal food intake cycles using the inertial signals (acceleration and orientation velocity) from an off-the-shelf smartwatch. We use five specific wrist micromovements to model the series of actions leading to and following an intake event (i.e., bite). Food intake detection is performed in two steps. In the first step, we process windows of raw sensor streams and estimate their micromovement probability distributions by means of a convolutional neural network. In the second step, we use a long short-term memory network to capture the temporal evolution and classify sequences of windows as food intake cycles. Evaluation is performed using a challenging dataset of 21 meals from 12 subjects. In our experiments, we compare the performance of our algorithm against three state-of-the-art approaches, where our approach achieves the highest F1 detection score (0.913 in the leave-one-subject-out experiment). The dataset used in the experiments is available at https://mug.ee.auth.gr/intake-cycle-detection/.

The structure of the cumulative food intake (CFI) curve has been associated with obesity and eating disorders. Scales that record the weight loss of a plate from which a subject eats food are used for capturing this curve; however, their measurements are contaminated by additive noise and are distorted by certain types of artifacts. This paper presents an algorithm for automatically processing continuous in-meal weight measurements in order to extract the clean CFI curve and in-meal eating indicators, such as total food intake and food intake rate. The algorithm relies on the representation of the weight-time series by a string of symbols that correspond to events such as bites or food additions. A context-free grammar is next used to model a meal as a sequence of such events. The selection of the most likely parse tree is finally used to determine the predicted eating sequence. The algorithm is evaluated on a dataset of 113 meals collected using the Mandometer, a scale that continuously samples plate weight during eating. We evaluate the effectiveness for seven indicators and for bite-instance detection. We compare our approach with three state-of-the-art algorithms, and achieve the lowest error rates for most indicators (24 g for total meal weight). The proposed algorithm extracts the parameters of the CFI curve automatically, eliminating the need for manual data processing, and thus facilitating large-scale studies of eating behavior.

Studying eating behaviours is important in the fields of eating disorders and obesity. However, the current methodologies of quantifying eating behaviour in a real-life setting are lacking, either in reliability (e.g. self-reports) or in scalability. In this descriptive study, we deployed previously evaluated laboratory-based methodologies in a Swedish high school, using the Mandometer®, together with video cameras and a dedicated mobile app in order to record eating behaviours in a sample of 41 students, 16–17 years old. Without disturbing the normal school life, we achieved a 97% data-retention rate, using methods fully accepted by the target population. The overall eating style of the students was similar across genders, with male students eating more than females, during lunches of similar lengths. While both groups took similar number of bites, males took larger bites across the meal. Interestingly, the recorded school lunches were as long as lunches recorded in a laboratory setting, which is characterised by the absence of social interactions and direct access to additional food. In conclusion, a larger scale use of our methods is feasible, but more hypotheses-based studies are needed to fully describe and evaluate the interactions between the school environment and the recorded eating behaviours.

In the context of dietary management, accurate monitoring of eating habits is receiving increased attention. Wearable sensors, combined with the connectivity and processing of modern smartphones, can be used to robustly extract objective and real-time measurements of human behavior. In particular, for the task of chewing detection, several approaches based on an in-ear microphone can be found in the literature, while other types of sensors have also been reported, such as strain sensors. In this paper, performed in the context of the SPLENDID project, we propose to combine an in-ear microphone with a photoplethysmography (PPG) sensor placed in the ear concha, in a new high accuracy and low sampling rate prototype chewing detection system. We propose a pipeline that initially processes each sensor signal separately, and then fuses both to perform the final detection. Features are extracted from each modality, and support vector machine (SVM) classifiers are used separately to perform snacking detection. Finally, we combine the SVM scores from both signals in a late-fusion scheme, which leads to increased eating detection accuracy. We evaluate the proposed eating monitoring system on a challenging, semifree living dataset of 14 subjects, which includes more than 60 h of audio and PPG signal recordings. Results show that fusing the audio and PPG signals significantly improves the effectiveness of eating event detection, achieving accuracy up to 0.938 and class-weighted accuracy up to 0.892.

This paper introduces fsLDA, a fast variational inference method for supervised LDA, which overcomes the computational limitations of the original supervised LDA and enables its application in large-scale video datasets. In addition to its scalability, our method also overcomes the drawbacks of standard, unsupervised LDA for video, including its focus on dominant but often irrelevant video information (e.g. background, camera motion). As a result, experiments in the UCF11 and UCF101 datasets show that our method consistently outperforms unsupervised LDA in every metric. Furthermore, analysis shows that class-relevant topics of fsLDA lead to sparse video representations and encapsulate high-level information corresponding to parts of video events, which we denote "micro-events".

This article presents a novel approach to training classifiers for concept detection using tags and a variant of Support Vector Machine that enables the usage of training weights per sample. Combined with an appropriate tag weighting mechanism, more relevant samples play a more important role in the calibration of the final concept-detector model. We propose a complete, automated framework that (i) calculates relevance scores for each image-concept pair based on image tags, (ii) transforms the scores into relevance probabilities and automatically annotates each image according to this probability, (iii) transforms either the relevance scores or the probabilities into appropriate training weights and finally, (iv) incorporates the training weights and the visual features into a Fuzzy Support Vector Machine classifier to build the concept-detector model. The framework can be applied to online public collections, by gathering a large pool of diverse images, and using the calculated probability to select a training set and the associated training weights. To evaluate our argument, we experiment on two large annotated datasets. Experiments highlight the retrieval effectiveness of the proposed approach. Furthermore, experiments with various levels of annotation error show that using weights derived from tags significantly increases the robustness of the resulting concept detectors.

In this paper we propose a novel approach to training noise-resilient concept detectors from clickthrough data collected by image search engines. We take advantage of the query logs to automatically produce concept detector training sets; these suffer though from label noise, i.e., erroneously assigned labels. We explore two alternative approaches for handling noisy training data at the classifier level by training concept detectors with two SVM variants: the Fuzzy SVM and the Power SVM. Experimental results on images collected from a professional image search engine indicate that 1) Fuzzy SVM outperforms both SVM and Power SVM and is the most effective approach towards handling label noise and 2) the performance gain of Fuzzy SVM compared to SVM increases progressively with the noise level in the training sets.

This paper presents the concept detector module developed for the VITALAS multimedia retrieval system. It outlines its architecture and major implementation aspects, including a set of procedures and tools that were used for the development of detectors for more than 500 concepts. The focus is on aspects that increase the system's scalability in terms of the number of concepts: collaborative concept definition and disambiguation, selection of small but sufficient training sets and efficient manual annotation. The proposed architecture uses cross-domain concept fusion to improve effectiveness and reduce the number of samples required for concept detector training. Two criteria are proposed for selecting the best predictors to use for fusion and their effectiveness is experimentally evaluated for 221 concepts on the TRECVID-2005 development set and 132 concepts on a set of images provided by the Belga news agency. In these experiments, cross-domain concept fusion performed better than early fusion for most concepts. Experiments with variable training set sizes also indicate that cross-domain concept fusion is more effective than early fusion when the training set size is small.

This paper proposes a methodology for modeling the process of semantic identification and controlling its complexity and accuracy of the results. Each semantic entity is defined in terms of lower level semantic entities and low level features that can be automatically extracted, while different membership degrees are assigned to each one of the entities participating in a definition, depending on their importance for the identification. By selecting only a subset of the features that are used to define a semantic entity both complexity and accuracy of the results are reduced. It is possible, however, to design the identification using the metrics introduced, so that satisfactory results are obtained, while complexity remains below some required limit

The process of automatic identification of high level semantic entities (e.g., objects, concepts or events) in multimedia documents requires processing by means of algorithms that are used for feature extraction, i.e. low level information needed for the analysis of these documents at a semantic level. This work copes with the high and often prohibitive computational complexity of this procedure. Emphasis is given to a dynamic scheme that allows for efficient distribution of the available computational resources in application. Scenarios that deal with the identification of multiple high level entities with strict simultaneous restrictions, such as real time applications

This paper addresses the problem of identifying semantic entities (e.g., events, objects, concepts etc.) in a particular environment (e.g., a multimedia document, a scene, a signal etc.) by means of an appropriately modelled semantic encyclopedia. Each semantic entity in the encyclopedia is defined in terms of other semantic entities as well as low level features, which we call syntactic entities, in a hierarchical scheme. Furthermore, a methodology is introduced, which can be used to evaluate the direct contribution of every syntactic feature of the document to the identification of semantic entities. This information allows us to estimate the quality of the result as well as the required computational cost of the search procedure and to balance between them. Our approach could be particularly important in real time and/or bulky search/indexing applications.