Consumer devices such as smartphones, wearables, and smart home assistants increasingly rely on machine learning models for personalized and secure services. Federated Learning (FL) enables decentralized training without sharing raw data by communicating model updates only, but faces challenges from heterogeneous clients and adversarial behavior that destabilize convergence. This paper introduces FedPIDAvg_tuned, a feedback-based FL aggregation that integrates a Proportional–Integral–Derivative (PID) controller with Bayesian Optimization to automatically stabilize the aggregation and improve robustness. Experiments on non-IID FEMNIST data demonstrate up to 4% higher accuracy and improved convergence stability compared to traditional FL aggregation, showing that adaptive control strengthens model robustness under adversarial conditions.

Statistical heterogeneity in Federated Learning (FL) causes gradient conflict, where updates from different clients oppose each other. Standard aggregation methods like FedAvg struggle in these non-IID settings, often degrading performance for minority clients and causing model divergence. To address this, we introduce FedHarmo, a model-agnostic aggregation strategy for personalized FL. FedHarmo operates by identifying and separating conflicting parameters at the client level. Each client computes a lightweight "conflict" score to detect parameters that are misaligned with the global update trend. These conflicting parameters are withheld from aggregation and updated only locally, allowing for client-specific specialization. Meanwhile, non-conflicting parameters are aggregated to build a robust global model. Our experiments demonstrate that FedHarmo improves peak accuracy for the majority of participating clients, which translates to a consistent 5% gain in the aggregated accuracy, without compromising convergence speed, thereby enhancing the end-user consumer experience.

Federated Learning (FL), an emerging decentralized Machine Learning (ML) approach, offers an effective avenue for training models on distributed data while safeguarding consumer privacy. Nevertheless, adversarial attacks initiated by malicious clients gaining access to confidential consumer data may deteriorate the learning process and performance. We develop the novel MADE-PI technique for the conventional FL that identifies and excludes malicious clients in the early model training phases. Our method incurs a novel local models’ preprocessing step performed before the aggregation procedure, in which we analyze the distribution of PI scores assigned to each client and exclude potentially poisoned data sources. We introduce Proportional – Integral (PI) scores as distance-based detection metrics inspired by PID control that consider the current distance between clients and the history of distances, through the proportional and integral terms respectively. Our empirical study on three distinct consumer application FL datasets, which represent use-cases from the intelligent transportation, handwriting recognition and medical scenarios, verifies that our approach is successful in identification and removal of compromised clients, mitigating threats to consumer model integrity and the application performance. The possible gain in learning effectiveness and efficiency against conventional methods is evaluated and analyzed.

In this research, we address a critical vulnerability of Federated Learning (FL) in practical applications: their susceptibility to adversarial attacks from malicious participants, which can severely degrade a global model’s performance. We propose MADE-PI, a novel defense mechanism designed to identify and exclude malicious clients early in the training process. Our method introduces a pre-aggregation step where the server evaluates each client update using a Proportional-Integral (PI) score. This score is a unique distance-based metric inspired by the PID controller, where a proportional term P measures the immediate deviation of a client’s update, and an integral term I tracks its behavior over time. By analyzing the distribution of these scores, our technique effectively flags and removes malicious contributions. We validate our approach through an empirical study on three datasets representing various applications of FL, including intelligent transportation, medical image analysis, and handwriting recognition. We concentrate on data poisoning as a highly relevant class of attacks. The results demonstrate that MADE-PI outperforms conventional defenses by providing superior malicious client detection and improving both learning efficiency and final model accuracy.

Machine Learning (ML) systems are integrated with other parts of modern cyberinfrastructure (CI), which forms the novel ML with Integrated Network (MLIN) structure. Various security tools are already employed in CI to detect malicious attacks. However, their operation is limited to a certain CI part or MLIN component without considering their integration and the ML end performance as the major indicator. We design and implement Knowledge Integration System Service (KISS) that introduces a novel approach to detect and classify attacks into adversarial attacks against ML end system (A-Attacks), and against base MLIN infrastructure (B-Attacks). Unlike traditional security mechanisms, KISS examines how attacks on individual MLIN components impact the overall ML end system performance, and extracts and integrates knowledge from each MLIN CI component to better distinguish between the attack types. As our major contributions, we (1) investigate the effects of A- and B-Attacks on ML systems, actively used in industry. We (2) develop the KISS prototype and verify it in practice. We demonstrate how KISS can be integrated into already established MLIN CI and combined with the traditional security tools. Our experiments verify that KISS improves attack detection and their initial classification between A- and B-Attacks in MLIN operational setups.

In this paper, we present an innovative approach to automate key event detection and highlights generation from soccer match videostreams that allows to improve accuracy and reliability, as well as to reduce data consumption and training time. Our method segments the videostream into distinct frames based on camera angles and activities, and integrates intelligent video analytics with additional visual information provided by broadcasters. As our major novelty in comparison to other intelligent soccer video analysis approaches, we deploy a Multi-Class Image Classifier to segment the video into wide-angle overviews, close-ups, and in-game replays, which allows us to improve the event detection performance and the quality of generated highlights. As another major novelty, we leverage YOLOv8 to detect events such as bookings, substitutions, and goals based on the additional information portrayed by broadcasters during the game. We evaluate our approach and demonstrate its advantage, when the additional information from broadcasters is available, against existing ones that analyze only the actions happening in the scene itself, such as the players' current positions and their actions between the frames. We evaluate our pipeline on a real soccer game recording and compare the highlights it generates with the official highlights provided by the broadcaster. Our pipeline demonstrates ample performance and is able to detect all key events in the game.

In this paper, we investigate the robustness of the ML end application performance to network Quality of Service (QoS) degradation, and the ways to improve it. We introduce a novel system approach to define the Machine Learning (ML) with Integrated Networks (MLINs) and describe how ML end performance can be employed to adjust network hyperparameters in order to prevent system Data Quality decrease. We investigate the interrelations between the network QoS degradation during the data transmission and ML image classification performance. We demonstrate how the studied interrelationships can be employed to produce recommendations on network adjustment in order to improve MLIN robustness. In particular, we propose an example of MLIN feedback system design that employs ML end performance as the major indicator to produce recommendations on network hyperparameters adjustment aimed at improving MLIN robustness.

The evolution of the Machine Learning (ML) has led to the emergence of Transfer Learning (TL) approach, which allows reusing pretrained industrial ML systems after their input data values or even an application domain has changed. In this paper, we investigate the TL process and its impact on the performance of ML-end systems with integrated network facilities. Especially, we focus on ML-systems designed for the classification of image media-files, transmitted over a network. Packet loss in a network is considered as a major input Data Quality (DQ) deterioration factor that can result in ML system classification performance degradation after pretraining on good inputs. To investigate the typical industrial TL process, we study the relationships between the ML model's last layer weights, hyperparameters, and classification performance throughout the retraining process. In addition, we conduct an empirical study to evaluate how the TL affects ML model performance in real application scenarios. For our experiments, we employ real image media-files, and transmit them over a real wireless network with inherent data losses for a classification on a remote ML-end system. According to our results, retraining ML models on corrupted data allows to enhance their robustness to a DQ degradation in the considered image classification scenarios. However, DQ influence on the ML system performance may vary depending on the data and system types.

Nowadays, the number of aerial unmanned vehicles (UAVs) is growing at a tremendous speed, as well as its technology. Therefore, it is essential to follow this growth with increasingly robust algorithms to be possible to exist cooperation between robots autonomously. One of the major events currently being developed in autonomous cooperation is relatively terrain classification where, this classification, is mainly important for emergency landings, mapping and decision making. This paper presents a robust computer vision system to sort terrain types using two main algorithms: Particle Swarm Optimization (PSO) and Gray-Level Co-Occurrence Matrix (GLCM). In addition to these two algorithms, a neural network was designed with the aim of increasing the probability of success of the proposed system. In order to evaluate this article, the system is validated using videos acquired onboard of a UAV with a RGB camera.