Available at: https://digitalcommons.calpoly.edu/theses/2994
Date of Award
4-2025
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Joydeep Mukherjee
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
The growing complexity of cloud-native microservices has intensified the need for robust anomaly detection mechanisms, particularly for misconfiguration-induced failures that often manifest subtly and propagate across service boundaries. This thesis proposes a comparative framework for runtime anomaly detection by transforming system metrics into frequency-domain representations—spectrograms using Short-Time Fourier Transform and scalograms using Continuous Wavelet Transform—and classifying them with Convolutional Neural Networks (CNNs). Evaluated on DeathStarBench, a production-grade microservices benchmark with injected anomalies in the Reservation and Recommendation services, the framework also includes baseline Long Short-Term Memory (LSTM) models trained on raw time-series features. Stratified 5-fold cross-validation shows that CNNs trained on scalograms outperform all other models, achieving 98.54% accuracy, 98.53% F1-score, and a Matthews Correlation Coefficient (MCC) of 0.9628, compared to 84.76% accuracy and 85.10% F1-score for the best LSTM model (128×8). Scalograms also deliver superior class-wise F1-scores for both anomalies—0.9451 for A1 and 0.9276 for A2—highlighting their strength in capturing bursty and slowly varying patterns often missed by fixed-window spectrograms and temporal baselines. These findings support the hypothesis that wavelet-based representations offer enhanced sensitivity and robustness for detecting misconfigurations in dynamic, resource-constrained microservice environments, paving the way for future research in multi-resolution anomaly detection and real-time telemetry analysis.