Date Approved

5-26-2026

Graduate Degree Type

Thesis

Degree Name

Cybersecurity (M.S.)

Degree Program

School of Computing and Information Systems

First Advisor

Samah Mansour

Academic Year

2025/2026

Abstract

Highly pathogenic avian influenza outbreaks, respiratory disease, heat stress, and silent equipment failures share one operational reality: they are detected too late because today’s poultry-health workflow is reactive, manual, and dependent on producers volunteering commercially sensitive data. This thesis presents a wake-on-anomaly federated architecture that addresses both the detection-latency problem and the privacy–adoption deadlock that has so far prevented cross-farm collaboration. The architecture is organized in two tiers. Tier 1 is a lightweight LSTM autoencoder that continuously screens four routine telemetry channels (water, feed, house temperature, activity proxy) and emits a per-window reconstruction-error score. A debounced k-of-m trigger with cooldown converts windows into events; the operating point is chosen by a three-objective Pareto sweep over false alerts, detection delay, and missed events. Tier 2 wakes only when Tier 1 fires and runs two on-farm modules: an audio classifier trained on 346 real broiler vocalizations using 86-dimensional MFCC+∆ + ∆∆+spectral features, and a flock-level visual symptom reporter based on Farnebäck optical flow and MOG2 with day-of-age normalization. Results are fused into an explainable farmer report. Cross-farm training uses federated learning (FedAvg, FedProx) on top of an exact federated StandardScaler that ships only (sum,sum2,n) per feature, so no raw telemetry leaves any farm. The system is evaluated on a literature-anchored synthetic testbed of five non-IID farms with biologically calibrated baselines (Gompertz growth, Aviagen Ross 308 / Cobb 500 intake tables) and five injected anomaly types. Tier 1 reaches AUPRC= 0.68, exceeding moving-average (0.66), CUSUM (0.31), and age-stratified (0.64) baselines. Federation reduces mean false alerts/day from 0.876 (local-only) to 0.404 (FedAvg, −54%). The personalized FedAvg backbone yields 0.487 FA/day before Tier 2; OR-rule fusion drives false alerts to 0.000/day while retaining 76% true-event recall. The audio classifier reaches macro-F1= 0.94 on a file-level split.

Download

Share

COinS