October 24 ~ 25, 2026, Vienna, Austria
Meir Ariel School of Electrical and Computer Engineering, Faculty of Engineering, Tel Aviv University, Tel Aviv, Israel
We present a novel post-quantum cryptographic framework based on masked high-memory convolutional codes in the Niederreiter paradigm. The public key is derived from a high memory polynomial parity check matrix augmented by dense random masking and additional invertible linear transformations. The resulting matrix exhibits strong random like properties, effectively concealing its algebraic structure and resisting known structural and algebraic attacks. Legitimate recipients retain polynomial time decoding capability, whereas adversaries appear to be limited to generic information set decoding with exponential complexity. Under standard cryptanalytic estimates, the proposed construction achieves security margins exceeding those of classical McEliece and Niederreiter-based systems by factors greater than 2²⁰⁰. Beyond its enhanced security profile, the construction offers considerable design flexibility, supporting arbitrary plaintext lengths, linear-time decryption, and a uniform per-bit computational cost, enabling efficient scalability to very long messages. Practical implementation is facilitated by a parallel array of syndrome trellis decoders, dynamically instantiated according to candidate syndrome values, enabling efficient hardware and software implementations with high throughput. Overall, the proposed scheme is a promising candidate for robust, scalable, quantum-resistant public-key cryptography.
Code based cryptography, post quantum cryptography, convolutional codes, syndrome trellis.
Xintong (Tony) Jiang1, David T. Garcia2, 1Chino Hills High School, Chino Hills, CA 91709 2University of California, Los Angeles, CA, 90095
Multi-cat households often manage several litter boxes that require frequent manual inspection, a burden that grows with the number of cats in the home and that existing fully automated litter boxes address only at a cost that is prohibitive for many owners needing multiple units. This paper proposes LitterCheck, an affordable litter box monitoring device that pairs a Raspberry Pi, a YOLOv8-based camera system, and a BME680 volatile organic compound (VOC) gas sensor to notify owners when a litter box needs cleaning, without an automated cleaning mechanism. Development required resolving compatibility issues between the detection model and the Raspberry Pi's software environment, integrating a third-party camera module, and iterating on the physical enclosure across multiple board and camera changes. A bench session comparing the gas sensor's readings indoors and outdoors, evaluated using a two-sample decision-boundary procedure, revealed that the sensor's raw resistance reading does not stabilize quickly after power-on, and that the large difference observed between the two conditions was confounded with this unfinished warm-up rather than attributable to location alone. This finding indicates that a documented sensor settling period is necessary before a fixed VOC threshold can be deployed with confidence, and it is identified here as the primary direction for follow-up work. Compared to prior infrared- and ultrasonic-based designs, LitterCheck's camera-based detection avoids known ambient-temperature sensitivity and does not risk emitting a signal that could discourage the cat from using the litter box, positioning it as a low-cost, non-invasive complement to existing litter box monitoring approaches pending further validation of its gas-sensing threshold.
Litter box monitoring; Internet of Things; volatile organic compounds; computer vision; multi-cat households.
Viktoriia Vovchenko, Hochschule Bielefeld - Bielefeld University of Applied Sciences and Arts (HSBI), Germany
Instruction-guided image editing models have made photorealistic facial manipulation accessi- ble through natural language prompts, substantially lowering the barrier to misuse. These systems facilitate attacks such as identity theft, phishing, misinformation, and social engineering, while also challenging ex- isting forensic methods that were not designed for prompt-driven edits. XXX InstructFace-Edit (XIFE) [68] is a dataset of 36,000 edited facial image pairs at 1024 × 1024 resolution generated by four state-of-the- art instruction-guided image editing models (FLUX.1-Kontext-dev, FLUX.2-dev, Qwen-Image-Edit-2511, and Step1X-Edit-v1.2) across six semantically distinct edit types, curated using a multi-stage filtering pipeline that integrates structural, semantic, and multimodal large language model–based quality checks, with thresholds derived empirically from a human-annotated subset of 1,500 images. However, the pace at which new editing models are released routinely outstrips the construction cycle of any single dataset, so we extend XIFE with a fifth model, HunyuanImage 3.0, released shortly after the original dataset was com- pleted and which has since maintained – and continues to hold – a leading position on public image-editing leaderboards. To curate this extension, we introduce a new evaluation pipeline that combines human review with visual question answering (VQA), set it against the original VIEScore-based pipeline, and explain why we find the VQA-based workflow preferable for this specific case. This extension yields 17,868 additional edited pairs, bringing the dataset to a total of 53,868 facial image pairs across five instruction-guided editing models and six semantically distinct edit types, of which 26,205 are annotated as high quality. The full dataset, including all pairs with quality labels, is publicly available (URLs withheld for review), supporting future research on training, evaluating, and stress-testing detectors for instruction-guided facial manipulation
Synthetic face detection, Media forensics, Generative image editing, Dataset curation, Vision- language evaluation.
Chuluundorj Begz University of the Humanities, Mongolia
Morphological complexity and word order flexibility present a specific object of analysis in developing language models. Syntactic and morphological features on multilingual word encoding serve as a basis to analyze how linguistic typology shapes multilingual embeddings. Language-specific patterns in embedding space present key factors influencing multilingual embedding structures in developing cross-lingual models.
Positional encoding, dot product, word embedding,associative network, attention transformer,graph, vector model, sinusoidal function.
Prince Ukobasi Frank Faculty of Law, University of Uyo, Nigeria
Large language models generate fluent legal text but frequently produce fabricated legal authorities. Retrieval-Augmented Generation (RAG) is widely proposed as a remedy. This paper systematically reviews the empirical literature on LLM legal hallucination and RAG-based legal question answering and shows that evidence is concentrated in the United States, the EuropeanUnion, China, Korea, Canada, India, and Ukraine, with no located peer-reviewed evaluationof RAG-grounded legal question answering on Nigerian, or more broadly African, primary law. Using Nigerian law as a case study, we propose a fully specified evaluation framework- benchmark taxonomy, ground-truth protocol, RAG pipeline design, blinded evaluation rubric, andfailure-mode taxonomy - intended to make future empirical work on this gap reproducible. Thisis a review and methodology paper; it reports no new empirical results.
Retrieval-augmented generation, legal hallucination, low-resource NLP, Nigerian law, LLMevaluation.
Joseph T. Mathew1and Ajila S. Babu2 1,2Department of Computer Science and Engineering, Amrita Vishwa Vidyapeetham, Amritapuri Campus, Kollam, Kerala, India
Autonomous vehicle navigation requires robust perception and downstream control. Traditional sequential architectures struggle to evaluate how visual perception errors affect vehicle safety. We propose an integrated perception-to-control framework mapping visual spatial data into a 12- dimensional state vector for a Deep Q-Network (DQN) controller. To mitigate urban occlusion, a fine-tuned DenseNet-121 appearance extractor trained with triplet margin loss is integrated into DeepSORT. On 15 KITTI benchmark sequences, DenseNet-DeepSORT achieves 84.6% MOTA and reduces identity switches by 66.2%. Spatial tracklets from YOLOv8 and DenseNetDeepSORT are projected into 3D ego-centric coordinates to feed the DQN controller in a kinematic simulator. A deterministic safety override layer using kinematic braking constraints prevents unsafe policy actions caused by state-estimation noise. Across 1,000 simulated episodes, perception noise increases unconstrained DQN collisions from 1.58% to 15.8%, whereas our safety layer cuts collisions to 3.2% with a 96.8% driving success rate.
Autonomous Driving, Deep Reinforcement Learning, Multi-Object Tracking, DeepSORT, DenseNet121.
Dean Chiu1, David T. Garcia2, 1Ayala High School, 14255 Peyton Dr, Chino Hills, CA 91709 2University of California, Los Angeles, CA, 90095
As people grow, they acquire practical life skills that matter less for any single task than for how they are used across everyday life. Life Skills Game teaches several of these skills — cooking, internet safety, and home security — through interactive, objective-driven play. The cooking game teaches players to read a recipe, turn fridge ingredients into a finished meal, and clean up afterward. The internet-safety game places the player inside a virtual computer to find their mother's Gmail while avoiding scam emails that try to hinder their progress, presenting online safety in a realistic scenario where a careless player can be scammed or even hacked. A third module addresses at-home security. Building the game raised two main challenges: deciding which objectives each module should contain — which dishes, scams, or security threats to include — and correctly attaching code to the game's models so that they behave as intended, which took repeated iteration and in-game testing before the interactions worked reliably. Two design insights shaped the experiments that follow: adding a cybersecurity setup step before the computer game teaches players to secure a device before using it, and adding time management to the kitchen game makes cooking more realistic and challenging by asking players to work under pressure. The broader goal is for students who lack a foundation in these skills to gain enough familiarity, and inspiration, to use them confidently later in life.
Educational, Beneficial, Security, Cooking.
Ahmet Çay, Barış Bayram, Ayşe Dilara Türkmen, and Alaeddin Türkmen Data Science, Hepsijet, İstanbul, Turkey
Last-mile delivery platforms increasingly require automated verification that a courier-submitted proof-of-delivery photo actually shows a package left at the door, while withholding photos that expose a bystander’s or resident’s face for privacy reasons. We frame this as a joint two-class object detection problem (package, face) and study it end-to-end on a real-world delivery photo dataset with a severe, naturally occurring class imbalance: package instances outnumber face instances by roughly 90 to 1, since most delivery photos simply do not contain a visible person. We conduct a systematic Bayesian hyperparameter search over ten YOLO variants spanning two model generations (YOLO11 and YOLO26, each in n/s/m/l/x sizes), evaluating every resulting model against an independent, held-out set of real-world ground-truth photographs rather than a validation split drawn from the training distribution. Because the detector must run on-device on courier Android phones, we treat model size and inference speed as firstclass constraints alongside accuracy, favoring the smallest model that meets our accuracy bar rather than the single best-scoring one. We find that model scale alone does not reliably fix face detection under class imbalance: the smallest model fails to detect any face at all, while the best package-detection model is not the best face-detection model. To address face-class scarcity directly, we source supplementary face imagery from a large public face dataset, auto-labeled with a pretrained face detector, and test two augmentation strategies. Naively combining the full public dataset inverts the imbalance in the opposite direction and, despite leaving package performance unchanged, substantially degrades face-detection performance due to domain shift: crowd, sports, and stage scenes look nothing like a single person standing near a doorway. In contrast, curating a small, domain-matched subset of close-range single- and couple-face images improves face-detection F1 to its best result across every configuration tested, at negligible cost to package detection. We additionally show that a confidence threshold tuned on a validation split containing augmented data does not transfer to the real-world evaluation set, underscoring the necessity of deployment-matched validation for operating-point selection. The resulting system has been deployed in production nationwide across Turkey and remains under active monitoring and extension. Our results argue that for severely imbalanced detection problems, the domain fidelity of minority-class augmentation matters far more than its raw volume. .
object detection, YOLO, class imbalance, data augmentation, privacy-preserving computer vision, mobile deployment.
Barı¸s Bayram, Alaeddin T¨urkmen, Ahmet C¸ ay, and Ay¸se Dilara T¨urkmen Data Science, Hepsijet, ˙Istanbul, Turkey
Automated analysis of customer feedback is dominated by high-frequency complaint categories, while categories of operational importance are often rare, and methods that rank topics by prevalence deprioritise exactly those categories whose incidence later increases. We address the prospective form of the problem: at each point in time, using only the data available then, which rare categories should be flagged as growing. Categories are assigned by an open-weight large language model under zero-shot prompting, so that membership does not depend on corpus frequency, and a rate-based alarm is evaluated walk-forward over 3.3 million logistics reviews collected between 2019 and 2026. Against five baselines on identical decision points, the rule recovers all three growth events the corpus contains while raising 6.0 to 12.1 times fewer alarms than two non-parametric criteria; a sequential change-point baseline matches it. We further find that detection latency is governed by the expected count per period rather than by bin width, and that chronologically batched labelling induces spurious temporal variation in labelling rates. .
emerging topic detection, burst detection, topic detection and tracking, rare category detection, early warning, multi-label classification, zero-shot classification, customer feedback analysis .
Kadarla Ashritha¹, Shenishetty Varshitha¹, Anupoju Jyothi Vishal¹, Computer Science and Engineering, Vardhaman College of Engineering Kacharam, Shamshabad, 501218, Hyderabad, Telangana, India Hari Babu², Assistant Professor, Computer Science and Engineering, Vardhaman College of Engineering Kacharam, Shamshabad, 501218, Hyderabad, Telangana, India
A web application penetration testing toolkit is essential for protecting modern web systems. It combines automated and manual testing methods to identify vulnerabilities like SQL injection, Cross-Site Scripting(XSS), Cross-Site Request Forgery (CSRF), security misconfigurations, and session exploitation, as notedin the OWASP Top 10. Multi-user authentication, a responsive glassmorphism dashboard, HTTP header analysis, path enumeration, and payload injection (XSS reflection, SQLi tautologies) are all included inthe system. It takes 15 seconds to scan targets and categorize results based on severity (CRITICAL/HIGH/MEDIUM/INFO) when it is deployed on Render. Testing on benchmark websites (testphp.vulnweb.com) yielded 85% detection accuracy and 20% false positives. By bridging the gapbetween academic learning and production use, the toolkit allows SMEs and students to conduct OWASP-compliant assessments without the requirement for commercial products. .
Web Penetration Testing, OWASP Top 10, SQL Injection, XSS Detection, Vulnerability Scanner.
Walid Ben Ali & Noora Fetais Qatar University, Qatar
Diffusion models that generate network-flow records as images of their features have been proposed for augmenting rare attack classes. We ask whether the layout of those images matters. On UNSW-NB15, we compared nine ways of building training data: real records only, random oversampling, SMOTENC, an adapted TabDiff, and one class-conditional U-Net diffusion design with the canonical layout, three random layouts or a training-fitted IGTD layout. Over five end-to-end repetitions, a frozen XGBoost detector was scored by mean F1 over Analysis, Backdoor, Shellcode and Worms on the official test file. Against a margin of 0.0211, IGTD was practically equivalent to the canonical layout (+0.0026; 95% interval −0.0103 to +0.0155) and to the random-layout mean (+0.0008; −0.0124 to +0.0140). The three comparisons with real-only training were inconclusive. Many accepted U-Net rows held out-of-range values, and a post-hoc check that rejected them left equivalence with the canonical layout unconfirmed. .
Intrusion detection, Diffusion models, Data augmentation, Feature layout, Class imbalance.
Ana Landim1, Ana Karina Moreira2, António Raimundo1, Camila Monteiro2, Dominika Swolkien2, Erik Daskalyuk1, Etelvino Garcia2, Miguel Sales Dias1, Leandra Lopes2, Ricardo Pereira1, 1University Institute of Lisbon (ISCTE-IUL), ISTAR, Portugal, 2University of Cape Verde (Uni-CV), CiHU, Cape Verde
Cape Verdean Creole (Santiago variety, kea-ST) is a low-resource language lacking parallel corpora and translation tools. This paper presents the development of multilingual neural machine translation models between kea-ST and European Portuguese (pt-PT), developed by fine-tuning pretrained mBART-50, M2M-100, and NLLB-200 models. As part of the ORAL – kriOl(u) laRge lAnguage modeLs project, co-funded by the Organization of Ibero-American States (OEI) and developed in collaboration between ISTAR-Iscte and the University of Cape Verde (Uni-CV), a validated bilingual corpus of 16,727 sentence pairs was compiled and reviewed by language specialists via the KeaColab platform. Using the CRISP-DM methodology, models were evaluated with BLEU, chrF, and ROUGE-L metrics plus qualitative linguist review. mBART-50 achieved the best performance (73.25 BLEU, 85.28 chrF, 86.78 ROUGE-L), showing strong fidelity and fluency. Results confirm that fine-tuning pretrained multilingual models effectively improves machine translation for low-resource languages like Cape Verdean Creole, advancing NLP resources for underrepresented languages.
Neural Machine Translation, Cape Verdean Creole, European Portuguese, Multilingual Large Language Models, Generative Artificial Intelligence.
Chuluundorj Begz, University of the Humanities, Mongolia
Morphological complexity and word order flexibility present a specific object of analysis in developing language models. Syntactic and morphological features on multilingual word encoding serve as a basis to analyze how linguistic typology shapes multilingual embeddings. Language-specific patterns in embedding space present key factors influencing multilingual embedding structures in developing cross-lingual models.
Positional encoding, dot product, word embedding,associative network, attention transformer,graph, vector model, sinusoidal function.
Jiajin Li1, Ang Li2 1Marlborough School, 250 S Rossmore Ave, Los Angeles, CA 90004 2California State University Long Beach, 1250 Bellflower Blvd, Long Beach, CA 90840
Lameness affects horses on half of U.S. horse operations each year, yet mild cases are difficult to detect, and even experienced veterinarians often disagree about subtle lameness. This paper presents HorsiSense, a mobile application paired with a wearable sensor that records a horse's orientation, location, and speed and turns genuine measurements into records and AI insights. The system combines a Particle Boron microcontroller with an inertial measurement unit and GPS module, a React Native application that pairs with the device over Bluetooth Low Energy in one step, and an AI Insights module that only presents findings supported by real recorded data. Challenges including sensor detection, cross-platform Bluetooth permissions, and AI hallucination were addressed through read-based device scanning, version-specific permission handling, and data-gated insights. Experiments showed a mean orientation error of about 1.2 degrees and reliable Bluetooth delivery to 10 metres, while revealing indoor magnetic interference and body shielding as limitations. HorsiSense offers owners an accessible, transparent way to monitor movement between veterinary visits.
Equine Lameness, Inertial Measurement Unit, Bluetooth Low Energy, Mobile Application, Artificial Intelligence.