As IoT networks become more complex, so too does their vulnerability to various cyber threats and attacks, making the need for strong IDS an absolute must. This paper investigates IDS machine learning models based on both a single data set as well as cross-boundary datasets. Supervised, unsupervised, and hybrid methods (Random Forest, Isolation Forest, Autoencoder, etc.), as well as their combinations, were evaluated using UNSW-NB15 and CICIDS2017 datasets. Where most studies evaluate models using training and testing sets within the same data set, this study primarily evaluated their ability to generalize to a different dataset. Based on experimental results, supervised models performed well for training and testing within the same data set; however, their ability to generalize was drastically reduced when training and testing were performed across datasets (~90% F1 score for r. Forest vs. 52% F1 score for r. Forest when trained/tested on same dataset). Conversely, while unsupervised models lack accuracy when training/testing on controlled data sets, they do show relatively consistent performance when comparing their performance across different data distributions due to dataset distribution biases. Based upon feature distribution comparisons among data distributions, it appears that differences in dataset distributions have the greatest effect on the generalization of current IDS models. This study concluded that there are limitations to the current state-of-the-art methods of IDS and that in any design of an IDS for a real-world IOT scenario; dataset distribution shifts must be accounted for when providing for reliable INS
Fingerprint recognition systems often struggle when the fingerprint image is of quality. Many systems deal with this by applying the enhancement method to every image regardless of how good or bad the image is. This one-size-fits-all approach may not work well when image quality changes a lot from one image to the other. This paper introduces a framework that first checks the quality of a fingerprint image and then picks a preprocessing method based on the quality tier: Poor, Medium or Good. A Random Forest classifier, trained using seven image-quality features was used to sort images into these three categories. On a test set of 75 subjects (600 images) that the classifier had never seen before it correctly classified 99.5% of the images. For each quality tier, we explored a set of 58 valid preprocessing sequences. These sequences were made up of five operations: normalization, denoising, contrast enhancement, ridge sharpening and segmentation. We picked the sequence that gave the Equal Error Rate (EER) on a separate tuning set. We then tested three conditions on the held-out evaluation set using the official FVC2004 matching protocol with 4,875 genuine and impostor pairs: no preprocessing (just raw images), one fixed preprocessing sequence applied to all images, and our adaptive quality-based pipeline. The adaptive pipeline performed best in the Poor-quality tier achieving an EER of 18.01% compared to 18.45% for images. However, in the medium and good tiers raw images did as well as or even better, than both the fixed and adaptive pipelines. We look into why this pattern might have happened and what it tells us about how to design preprocessing studies
Synthetic identity fraud (SIF) combines real and fabricated personal information to create new identities that pass onboarding checks, build credit over time, and disappear in a bust-out. Traditional rule-based models often miss these identities, and generative AI and deepfakes make them more realistic. This narrative review examines how researchers detect SIF using behavioral and network features. It covers generative-contrastive approaches to identity generation and detection, vulnerabilities in Know Your Customer (KYC) processes, temporal and transactional behavioral features, behavioral biometrics, and transaction network topology for finding fraud rings. The reviewed studies indicate that SIF rings leave coordinated temporal traces, that behavioral interaction data is difficult to forge, and that synthetic identities tend to form dense local clusters with few ties to the wider financial network. The review adds a summary table of feature families and notes that fraudsters can adapt to graph-based detection. Most evidence comes from single studies, common benchmarks are missing, and behavioral monitoring raises privacy and fairness questions. The article follows no formal search protocol. Future work should test detection methods under adversarial conditions and report results in a standard format.
Lost USB flash drives remain a steady source of personal-data incidents, and the organization that loses a drive usually cannot tell whether anyone has opened the files on it. We study whether honeytokens, which raise an alert when a file is opened, can remove part of that uncertainty. The paper makes three contributions. First, we analyze 300 publicly reported loss incidents from 2005 to 2024 and find that schools and universities (97 cases) and hospitals (76 cases) account for more than half of them. Second, we test ten Canarytokens token types on macOS, Windows 10 and Windows 11, with and without security software: Word and Excel tokens alerted in every environment, whereas PDF and executable tokens were unreliable or were blocked. Third, because the public service can only issue new template files, we build a self-hosted system that embeds a token into a user's existing Excel workbook and collects every alert in a searchable log. In a preliminary evaluation with four participants and 16 file openings, every opening raised an alert whose timestamp matched the recorded opening time to the minute. HTTP tokens located the opener to the correct prefecture or municipality in 6 of 8 openings; DNS tokens did so in none. We discuss what these results do and do not show, and what remains to be evaluated
— Identifying animals manually in camera-trap images, underwater images and sound recordings is slow and costly, but recording species is the first step in protecting animals. Deep learning can identify species automatically, but studies use different data and different scores, so it is hard to know which approach works best. This review examines ten deep-learning studies from 2023 to 2025 and one field dataset. It compares single-model approaches, such as convolutional neural networks and transformers, with hybrid approaches that join two or more models or kinds of data. To keep the comparison fair, hybrid and single models are compared only within the same study, on the same data and with the same measure, such as accuracy or F1-score. Hybrids scored higher than single models in every study that tested both, but in four of the five studies the gain was only 0.41 to 4.2 points, and only one study reported the extra cost. The type of data mattered more than the type of model. The same models fell from a mean average precision of 0.71 on clean recordings to 0.311 on field soundscapes, and the highest accuracies came from studies whose data splits may let near-copies of examples appear in both training and test sets. The remaining problems are missing spread between runs, missing costs, data leakage, weak results on rare species and the drop on field data. Future studies should report repeated runs and costs, split data by recording or site, and test on shared field datasets.
