This community contains the ePrints and eTheses produced by Computing Science and Mathematics staff and students. http://hdl.handle.net/1893/35 2026-10-24T08:14:33Z 2026-10-24T08:14:33Z Takahashi, Osamu Cuyt, Annie O'Hare, Anthony Lee, Wen-shin Ferrara, Massimiilinao http://hdl.handle.net/1893/38410 2026-10-23T00:01:15Z 2026-06-15T00:00:00Z

Title: An Explainable Estimation Model for Financial Time Series Author(s): Takahashi, Osamu; Cuyt, Annie; O'Hare, Anthony; Lee, Wen-shin; Ferrara, Massimiilinao Editor(s): Vantaggi, Barbara; Coletti, Giulianella; Denoeux, Thierry; Laurent, Anne; Petturiti, Davide; Miranda, Enrique; Medina, Jesus; Bouchon-Meunier, Bernadette; Yager, Ronald R Abstract: Recent developments in artificial intelligence have led to powerful machine learning approaches to model fluctuating dynamics in financial time series, even in the presence of non-stationary behavior. However, such methods often suffer from limited interpretability while uncertainty, arising from complex data fluctuations, remains difficult to quantify. We present a method that decomposes financial data into different frequency components and trends, thereby enabling explicit identification of cyclical patterns and long-term trends. Concurrently, noise components are separated from signal components. Its effectiveness and its high predictive accuracy have been demonstrated in the interpretation of some financial time series.

2026-06-15T00:00:00Z Hasan, Syed Azmul Hera, Mowmita Pervin Mim, Afroza Lona Jihad Hossain, M M Ali, Hazrat Bulbul, Mohammad Farhad http://hdl.handle.net/1893/38409 2026-10-23T00:00:36Z 2026-10-19T00:00:00Z

Title: Beyond Frame‐Level Accuracy: Analyzing Video‐Level Stability in Transformer‐Based Deepfake Forensics Author(s): Hasan, Syed Azmul; Hera, Mowmita Pervin; Mim, Afroza Lona; Jihad Hossain, M M; Ali, Hazrat; Bulbul, Mohammad Farhad Abstract: Deepfake videos generated by advanced generative models pose serious threats to digital security, misinformation control, and media forensics. Although numerous detection methods have been proposed, their reported performances are often based on heterogeneous preprocessing pipelines, inconsistent data splits, and model-specific training protocols, hindering fair architectural comparison. This work presents a controlled and reproducible comparative study of convolutional neural networks (ResNet-18 and EfficientNet-B0) and a Transformer-based architecture (ViT-B/16) within a unified, face-centric deepfake detection framework. All models are trained using identical preprocessing, augmentation, optimization settings, and data splits on Celeb-DF v2 and FaceForensics++, and results are validated across three independent random seeds with bootstrapped confidence intervals. Experimental results show that while CNN-based models perform competitively at the frame level, their performance degrades after video-level aggregation. In contrast, ViT-B/16 demonstrates more stable video-level performance, achieving 92.8% (95% CI: 91.7–94.0) video-level accuracy and an area under the curve (AUC) of 0.950 (95% CI: 0.934–0.964) on Celeb-DF v2 (3-seed bootstrapped). Cross-dataset evaluation across the same three random seeds reveals a large, asymmetric generalization gap (AUC 0.79 FF++ ⟶ Celeb-DF vs. 0.64 Celeb-DF ⟶ FF++), indicating that in-domain video-level stability does not transfer to unseen manipulation domains. These findings indicate that global self-attention improves robustness and video-level stability for deepfake detection within a given data domain, while highlighting the need for further work on cross-domain generalization.

2026-10-19T00:00:00Z Kumar, K Kolberg, M http://hdl.handle.net/1893/38404 2026-10-22T09:32:17Z 2026-07-22T00:00:00Z

Title: Hybrid Scalable Scheduling Reservation Protocol for Industrial IoT Networks Author(s): Kumar, K; Kolberg, M Abstract: In dense Wireless Sensor Networks (WSNs), communication fails due to multipath fading and external interferences (EI). In Industry 4.0 environments, low-power wireless standards such as ZigBee, Bluetooth Low Energy (BLE), WirelessHART, and ISA100.11a face challenges in terms of reliability. Recently, Time-Slotted Channel Hopping (TSCH) protocol has been adopted as a default mode in IEEE 802.15.4, providing resilience against EIs and multipath fading through time synchronization and channel hopping. Although WirelessHART also adopts TSCH, its centralized and non-IP-based architecture introduces significant control overhead, which limits scalability and seamless Internet integration. BLE, relying on distinct MAC and physical-layer mechanisms, remains susceptible to packet loss under harsh industrial conditions. Conversely, traditional IP-enabled solutions that do not employ TSCH often depend on primitive channel-access protocols, leading to internal collisions and reduced reliability in dense deployments. 6TiSCH (RFC 9030) standard extends TSCH by providing IPv6 connectivity through a lightweight 6LoWPAN stack and the 6top sublayer, enabling integration with the IPv6 protocol suite and improved architectural flexibility. Despite these advantages, network scalability remains an open challenge, as existing 6TiSCH Scheduling Functions (SFs) operating over the 6top sublayer impose trade-offs among reliability, latency, and energy consumption. In this paper, we present a Hybrid Scalable Scheduling Reservation Protocol (SSR), addressing important challenges with scalability, reliability, latency, and energy efficiency. The proposal, with its cake-slicing technique, network depth mapping, efficient traffic adaption, and congestion mitigation, shows potential for significant improvements over existing solutions like Minimal Scheduling Function (MSF) (RFC 9033). The experimental results show significant improvements over fixed hysteresis-based scheduling policies of MSF and demonstrate scalable performance using hundreds of IoT nodes within a subnet.

2026-07-22T00:00:00Z Dovhoshliubnyi, llia Soroush, Nima Sami, Ashkan Brownlee, Alexander http://hdl.handle.net/1893/38384 2026-10-08T13:00:11Z 2026-07-10T00:00:00Z

Title: What Do AI Agents Actually Change? An Empirical Taxonomy of Mutation Patterns in Performance-Improving Pull Requests Anonymous Anonymous Institution Author(s): Dovhoshliubnyi, llia; Soroush, Nima; Sami, Ashkan; Brownlee, Alexander Abstract: AI coding agents are black boxes: we cannot inspect how they generate code, but we can inspect what they change. This distinction matters for search-based software engineering (SBSE), where techniques such as genetic improvement depend on mutation operators that reflect how code is actually transformed. Of the 33,596 agent PRs in the AIDev dataset, less than 400 target performance (fewer than 1%), making each successful case a valuable window into otherwise opaque agent behaviour. We classify 1,254 performance-relevant diff hunks from 216 of these PRs, spanning five agent systems, against the 18-category syntactic mutation taxonomy of Even-Mendoza et al. (2025) using an LLM-as-a-judge pipeline. Three categories dominate: name modification (36.9%), object creation (26.3%), and type change (22.6%), a profile strikingly different from prior genetic improvement corpora where no change accounted for 84%. Each agent commits to a distinctive mutation vocabulary, and each performance strategy activates a largely disjoint category subset. Agent identity and target strategy are therefore informative priors that narrow the effective SBSE operator space from 18 categories to a handful per context. Replication package: https://anonymous.4open.science/r/ssbse-challenge-2026-710C/

2026-07-10T00:00:00Z