Research

Applied Research & Data Science

Software engineering, cybersecurity, and machine learning research grounded in data analysis, experimentation, and academic work at UW-Eau Claire.

Software EngineeringCybersecurityMachine LearningData AnalysisExperimental Design

2026

Download Paper (PDF)

165,769Readable images

Web-scraped Cheese Pics archive analyzed in the study.

1,824Raw categories

Cheese-named folders in the source archive.

1,238→1,033Label space

Raw recognition classes reduced through canonicalization.

43.8→49.0%Top-1

Reported Small baseline versus retrained canonical-label result.

Abstract

Filter, Recognize, Organize: A Four-Stage Pipeline for Fine-Grained Recognition on a Web-Scraped Archive of Cheese Images was submitted with Xiang Ma to the WACV 2027 Evaluations and Datasets Track.

The study analyzes 165,769 readable images across 1,824 cheese-named categories from the public Cheese Pics archive and proposes a four-stage workflow: CLIP-based packaging filtering, fine-grained ConvNeXt recognition, evaluation under an external cheese taxonomy, and name-based canonicalization of near-duplicate labels.

Under the reported experiments, canonicalizing the label space from 1,238 raw classes to 1,033 merged classes increased top-1 accuracy from 43.8% to 49.0%, while scaling ConvNeXt-Tiny to ConvNeXt-Small under the same 25-epoch budget increased top-1 by 0.2 percentage points.

The paper was not accepted in WACV 2027 Round 1. Peer review recognized the systematic analysis and reproducibility effort while identifying limitations in novelty, statistical support, Stage 1 validation, paper formatting, and an important inconsistency in the reported recognition-dataset counts.

Key Finding

Zero-shot CLIP reached 4.4% top-1 accuracy, while a trained linear probe on the same frozen CLIP image features reached 36.7%, suggesting that visual features were substantially more useful than rare-name text grounding in this setting.

Project Overview

The project grew from undergraduate cheese-classification research into a data-centric study of fine-grained recognition under realistic web-scraped noise. Rather than treating architecture choice as the only lever, the work examines how filtering, taxonomy design, and label canonicalization affect measured recognition performance.

A central motivation is that web search creates structurally noisy datasets: images may contain retail packaging instead of the target object, class frequencies are long-tailed, and separate folder names can refer to products that are visually indistinguishable at web-image resolution.

The paper was submitted on June 11, 2026 and later reviewed in WACV 2027 Conference Round 1. The final recommendation was reject, with the meta-review stating that the analysis was systematic but the evidence and contribution were not yet strong enough for acceptance.

Problem Statement

Fine-grained recognition becomes unreliable when a web-scraped archive contains off-target imagery, long-tailed class distributions, and inconsistent label granularity.

The study asks how much recognition performance changes when the dataset itself is reorganized, and whether those gains exceed the gains from a moderate increase in backbone capacity under the evaluated conditions.

It also examines whether errors that appear wrong at the exact cheese-name level remain semantically reasonable when predictions are evaluated under a coarser external taxonomy.

Dataset Discussion

The source archive contains 165,769 readable images across 1,824 cheese-named folders from the public Kaggle Cheese Pics dataset.

Stage 1 uses 208 hand-labeled examples, split nearly evenly between unpackaged cheese and packaging-dominant images, to train a CLIP-feature logistic-regression filter.

The filter retains 62,174 images, or 37.5% of the archive. The recognition stage uses 1,238 cheese categories with at least 20 retained images.

The submitted manuscript reports 43,489 images in the capped Stage 2 set but elsewhere reports a fixed split of 39,299 training, 8,388 validation, and 8,388 test images. Peer review correctly flagged that these counts cannot simultaneously describe the same capped recognition dataset; this must be reconciled before any future submission.

Methodology

Extracted 768-dimensional L2-normalized CLIP ViT-L/14 embeddings and trained a balanced logistic-regression probe to distinguish unpackaged cheese from packaging-dominant images.
Compared zero-shot CLIP, a frozen-feature CLIP linear probe, and end-to-end ImageNet-pretrained ConvNeXt fine-tuning for 1,238-way cheese-type recognition.
Fine-tuned ConvNeXt-Tiny and ConvNeXt-Small with AdamW, cosine scheduling, label smoothing, RandAugment, mixed precision, and a fixed stratified train/validation/test split.
Mapped exact-type predictions into an externally defined cheese-family taxonomy to evaluate hierarchical consistency without deriving the taxonomy from the recognition model itself.
Canonicalized near-duplicate class names using normalized names and external catalog anchors, reducing 1,238 raw categories to 1,033 canonical classes.
Compared raw-label recognition, backbone scaling, relabel-only canonical scoring, retraining on merged labels, and class-balanced loss under otherwise controlled settings.

Key Findings

Zero-shot CLIP reached 4.4% top-1 accuracy, while a trained linear probe on the same frozen CLIP image features reached 36.7%, suggesting that visual features were substantially more useful than rare-name text grounding in this setting.
ConvNeXt-Tiny reached 43.6% top-1 and ConvNeXt-Small reached 43.8% top-1 under the reported 25-epoch setup.
Rolling exact-type predictions into the external taxonomy produced 64.7% accuracy over eight cheese families and 71.7% over a coarser five-family grouping.
Canonical-label scoring increased top-1 from 43.8% to 48.6% without retraining, and retraining on the 1,033 merged classes reached 49.0%.
Class-balanced loss improved the minority-class objective but reduced overall top-1 accuracy to 43.8% under the reported test distribution.
The strongest supported conclusion is conditional: within the evaluated dataset, model scales, and training budget, label-space refinement produced a larger measured gain than the Tiny-to-Small backbone increase.

Challenges

The archive combines visual noise with semantic label noise, so improvements can come from changing the definition of the task as well as from improving the visual representation.
Rare proper-noun cheese names are difficult for zero-shot language grounding even when CLIP image features contain useful discriminative structure.
The 5.2-point canonicalization gain changes the label space from 1,238 to 1,033 classes, so it should not be interpreted as directly equivalent to a representation-learning gain on an unchanged task.
The main Tiny-versus-Small comparison uses one seed, one split, one GPU, and one capacity step, making the reported 0.2-point difference too small to treat as statistically established without variance estimates.
The Stage 1 packaging filter gates all downstream experiments but was not evaluated with held-out precision and recall in the submitted manuscript.
Peer review identified a mathematical inconsistency between the Stage 2 dataset-size table, the 40-image-per-class cap, and the reported train/validation/test counts.

Results

Stage 1 retained 62,174 of 165,769 images as usable cheese imagery, a 37.5% retention rate.
Stage 2 results were 4.4% top-1 for zero-shot CLIP, 36.7% for the frozen CLIP linear probe, 43.6% for ConvNeXt-Tiny, and 43.8% for ConvNeXt-Small.
Stage 3 achieved 64.7% accuracy under the eight-family taxonomy and 71.7% under the five-family grouping.
Stage 4 produced 48.6% top-1 when existing predictions were scored after canonical relabeling and 49.0% after retraining on 1,033 merged classes; top-5 reached 71.2%.
The WACV Round 1 reviews praised the understandable pipeline, controlled experimental setup, anti-circularity precautions, reproducibility, and practical data-curation message.
The submission was rejected after reviewers raised concerns about limited novelty, insufficient capacity comparisons, lack of variance estimates, missing held-out Stage 1 metrics, formatting compliance, and inconsistent dataset statistics.

Research Outcomes

Co-authored and submitted a six-page student research paper to the WACV 2027 Evaluations and Datasets Track.
Built an end-to-end four-stage research pipeline spanning data filtering, fine-grained recognition, hierarchical evaluation, and label-space canonicalization.
Ran controlled ConvNeXt, CLIP, label-merging, and class-balancing experiments on a large real-world web-scraped archive.
Produced quantitative and qualitative analyses including top-1/top-5 comparisons, a family confusion matrix, training curves, and near-twin error examples.
Received detailed external peer review that identified concrete methodological and reporting improvements for a stronger future version.

Skills Demonstrated

Computer vision
Data-centric machine learning
Fine-grained visual recognition
CLIP feature extraction
ConvNeXt fine-tuning
Dataset curation
Label-space design
Long-tailed recognition
Taxonomy design
Ablation studies
Scientific writing
Peer-review analysis
Reproducible experimentation

Technology Stack

PythonPyTorchOpenCLIPConvNeXtScikit-learnNumPyPandasMatplotlibKaggleGitGitHubRTX 4060

Future Work Opportunities

Reconcile and regenerate every dataset-count statistic from a single versioned split manifest so the recognition-set size, per-class cap, and train/validation/test totals are mathematically consistent.
Evaluate the Stage 1 packaging filter on a genuinely held-out labeled set and report precision, recall, F1, and threshold-dependent error analysis.
Run multiple seeds or repeated splits for the Tiny, Small, and merged-label experiments and report means, variance, and confidence intervals.
Broaden the capacity comparison with at least one additional model scale, training budget, or resolution before making stronger capacity-versus-curation claims.
Separate the definitional benefit of reducing the number of classes from any representational benefit obtained by retraining after canonicalization.
Validate the Stage 3 fallback taxonomy assignments and quantify sensitivity to errors in the CLIP text-prototype fallback.
Expand related work around data-centric AI, dataset curation, label-space refinement, and noisy fine-grained recognition.
Use the correct target-venue template and perform a publication-consistency audit across tables, figures, text, and generated statistics before resubmission elsewhere.

2026

27M+Packet records

Kitsune IoT network attack dataset scale.

115Features

Engineered behavioral traffic features.

5Time windows

Historical windows used for temporal behavior.

6Evaluation views

Accuracy, precision, recall, F1, ROC-AUC, and confusion matrix.

Abstract

This research was completed for IS 480: Advanced Business Analytics and investigated whether supervised machine learning could distinguish multiple IoT network attack types from engineered behavioral traffic statistics.

The work uses the Kitsune Network Attack dataset from the UCI Machine Learning Repository, which contains more than 27 million packet-level records from a commercial IP-based surveillance and IoT environment under controlled attack scenarios.

Each packet is represented by 115 numerical features extracted across five historical time windows, creating a classification problem that is both high-volume and security-sensitive.

Key Finding

Temporal behavioral features are a strong basis for separating attack categories because they encode how traffic changes over multiple time windows.

Project Overview

The project focused on multi-class attack detection instead of simple benign-versus-malicious classification. That framing is more useful for security analysis because different attack families imply different response paths.

The research emphasized data preparation, stratified sampling, model comparison, evaluation discipline, and feature-importance analysis across temporal network behavior.

Problem Statement

IoT devices are vulnerable to denial-of-service, man-in-the-middle, reconnaissance, and botnet activity, but traditional signature-based intrusion detection can struggle with behavioral or previously unseen threats.

The project asked whether temporal behavioral statistics could classify SYN DoS, ARP MitM, Mirai Botnet, SSDP Flood, and benign traffic in a way that could support practical detection workflows.

Dataset Discussion

Dataset scale is more than 27 million packet records, so the research uses stratified sampling to keep experiments computationally feasible while preserving class distribution.

The feature table contains 115 numerical behavioral traffic features. These are extracted across five historical time windows, giving the models information about recent and longer-running traffic behavior.

Methodology

Loaded and merged selected attack datasets, then structured labels for multi-class classification.
Applied stratified sampling to preserve class balance while reducing the computational load.
Scaled numerical features where required by the model family.
Evaluated Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting where feasible.
Compared models with accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix analysis.
Placed special emphasis on recall and F1-score because missed attacks are more costly in a security context.

Key Findings

Temporal behavioral features are a strong basis for separating attack categories because they encode how traffic changes over multiple time windows.
Multi-class classification is more informative for defenders than binary detection because each attack family has different operational implications.
Recall and F1-score deserve more attention than raw accuracy when evaluating security-oriented detection models.
Feature importance analysis can help translate model behavior into practical indicators of network abuse.

Challenges

The dataset size required careful sampling and memory-aware processing before model comparison could be practical.
Class imbalance made aggregate accuracy less informative than recall, F1-score, and class-level confusion matrix analysis.
Feature importance needed to be interpreted in terms of traffic behavior rather than treated as a purely mathematical ranking.

Results

The research established a repeatable model-evaluation workflow for large-scale IoT traffic classification using stratified sampling, numerical preprocessing, class-aware metrics, and confusion-matrix analysis.
The course research is complete, but no verified final accuracy value is retained in the source material used for this portfolio. The page therefore documents the validated methodology, scope, and evaluation criteria without inventing performance metrics.
The completed work is presented as a methodology- and security-analysis case study rather than as a benchmark claim.

Research Outcomes

Processed and sampled from a 27M+ packet IoT network dataset.
Built a multi-class detection framing for SYN DoS, ARP MitM, Mirai Botnet, SSDP Flood, and benign traffic.
Defined an evaluation approach using precision, recall, F1-score, ROC-AUC, and confusion matrices.
Connected feature importance work to cybersecurity interpretation.

Skills Demonstrated

Cybersecurity analysis
Supervised classification
Large dataset sampling
Feature engineering interpretation
Model evaluation
Security-focused metrics
Experimental design

Technology Stack

PythonScikit-learnPandasNumPyMachine LearningData MiningSecurity Analytics

Future Work Opportunities

Reproduce the original experiments in a fully versioned pipeline so final model metrics can be published with complete provenance.
Expand error analysis around attack categories that are most often confused.
Use feature-importance findings to explain which temporal traffic signals matter most for each attack type.
Compare classical models with modern anomaly-detection or representation-learning approaches while preserving interpretable security metrics.

2025

2Modalities

Natural-language captions and Minecraft gameplay visuals.

CNN+Model families

Convolutional and transformer-oriented approaches explored.

R@KEvaluation direction

Retrieval-oriented metrics are the natural next evaluation step.

PrototypeResearch stage

Completed exploratory academic research rather than a production benchmark.

Abstract

This project explored how natural-language descriptions can be associated with visual Minecraft gameplay content using deep learning and multimodal representation techniques.

The work centered on constructing paired gameplay-and-caption data, experimenting with convolutional and transformer-based model ideas, and framing the problem as retrieval rather than only conventional image classification.

The project remained an academic prototype, so the portfolio emphasizes the data pipeline, modeling direction, and research lessons rather than unsupported benchmark claims.

Key Finding

Retrieval is a better framing than fixed-label classification when the target is a free-form description of gameplay.

Project Overview

The core research question was whether semantically similar gameplay moments and text descriptions could be represented closely enough for caption-to-gameplay retrieval.

Minecraft provides a useful but difficult domain because visually similar scenes can have very different semantic meanings depending on actions, entities, inventory state, and surrounding context.

Problem Statement

Traditional image classification assumes a fixed label set, while free-form captions contain richer semantic information and may describe actions, environments, entities, or intent at the same time.

The project therefore treated the task as a multimodal representation problem: learn or evaluate representations that can connect a textual description with the most relevant gameplay sample.

Dataset Discussion

The prototype used Minecraft gameplay samples paired with human-written natural-language captions.

Dataset work included organizing gameplay media, pairing descriptions with samples, and preparing the content for model experimentation in notebook-based environments.

Methodology

Prepared paired gameplay and caption samples for multimodal experiments.
Prototyped convolutional approaches for learning visual representations.
Explored transformer and Hugging Face tooling for language-aware and multimodal model directions.
Used Google Colab for iterative experimentation and model prototyping.
Framed evaluation around retrieval quality and semantic correspondence rather than a single fixed-class accuracy score.

Key Findings

Retrieval is a better framing than fixed-label classification when the target is a free-form description of gameplay.
Dataset quality and caption consistency are critical because multimodal models learn relationships between both modalities rather than from images alone.
Transformer-based representations provide a natural path for connecting language semantics with learned visual features.
Small prototype datasets are useful for validating pipelines but are not sufficient for strong generalization claims.

Challenges

Gameplay scenes can look visually similar even when the player's action or semantic context differs significantly.
A small research dataset limits how well larger multimodal architectures can generalize.
Caption quality and consistency directly affect retrieval experiments because multiple descriptions may reasonably fit the same gameplay scene.

Results

Built an initial data pipeline connecting gameplay samples with human-authored captions.
Prototyped CNN- and transformer-oriented model directions using Python, TensorFlow, Hugging Face, and Google Colab.
Established a retrieval-oriented research framing that could be extended with stronger pretrained multimodal encoders and a larger dataset.

Research Outcomes

Designed a paired visual-language dataset workflow for Minecraft gameplay.
Explored both convolutional and transformer-based representation approaches.
Developed practical experience with multimodal learning, deep learning experimentation, and notebook-based research workflows.
Documented the limitations of small-domain multimodal datasets and the importance of retrieval-oriented evaluation.

Skills Demonstrated

Multimodal learning
Deep learning
Natural-language representations
Computer vision
Dataset preparation
Retrieval systems
Research prototyping

Technology Stack

PythonTensorFlowHugging FaceGoogle ColabDeep LearningComputer VisionNLP

Future Work Opportunities

Expand the paired gameplay-caption dataset with more varied actions, biomes, entities, and player states.
Evaluate pretrained vision-language models as embedding backbones.
Introduce quantitative retrieval metrics such as Recall@K and mean reciprocal rank.
Compare frame-based retrieval with short video-clip representations that preserve temporal context.

Recent applied research first, with methodology and outcomes expanded in context.

Abstract

Project Overview

Problem Statement

Dataset Discussion

Methodology

Key Findings

Challenges

Results

Research Outcomes

Skills Demonstrated

Technology Stack

Future Work Opportunities