Location
Havener Center, Miner Lounge / Wiese Atrium, 1:30pm-3:30pm
Start Date
4-2-2026 1:30 PM
End Date
4-2-2026 3:30 PM
Presentation Date
April 2, 2026; 1:30pm-3:30pm
Description
Thermal object detection systems are critical for safety sensitive applications due to their reliability under adverse conditions. However, existing robustness evaluations in thermal domain primarily focuses on physical or sensor-level perturbations, overlooking vulnerabilities from semantically realistic scene and object manipulations. We introduce Thermal Mirage, a generative framework that leverages GAN-guided diffusion to expose weaknesses in thermal detectors through controlled object and context level perturbations. Our approach learns class-conditional thermal priors via a GAN and uses diffusion to transform object appearances into ambiguous or low-saliency patterns. Simultaneously, a context module degrades scene conditions by simulating harsher night environments. Integrated with a detector-in-the-loop framework, our method generates realistic yet challenging inputs that reveal failure modes beyond traditional benchmarks. Experiments show significant performance degradation, while incorporating these scenarios into training improves robustness and generalization, highlighting the importance of semantic and context-aware evaluation in thermal perception.
Biography
Nuzaer Omar is a Kummer Doctoral Fellow and currently pursuing her Ph.D. in the Department of Computer Science at Missouri University of Science & Technology. Her research focuses on the development of robust and trustworthy ML/LLM systems, with a focus on black-box/ zero-access adversarial attacks. She is particularly interested in exploring how subtle perturbations affect model performance and devising strategies to mitigate these vulnerabilities. Her work contributes to advancing trustworthy and reliable AI systems in real-world applications
Meeting Name
2026 - Miners Solving for Tomorrow Research Conference
Department(s)
Computer Science
Document Type
Poster
Document Version
Final Version
File Type
event
Language(s)
English
Rights
© 2026 The Authors, All rights reserved
Included in
Thermal Mirage: Towards Robust Thermal Perception via GAN-Guided Diffusion
Havener Center, Miner Lounge / Wiese Atrium, 1:30pm-3:30pm
Thermal object detection systems are critical for safety sensitive applications due to their reliability under adverse conditions. However, existing robustness evaluations in thermal domain primarily focuses on physical or sensor-level perturbations, overlooking vulnerabilities from semantically realistic scene and object manipulations. We introduce Thermal Mirage, a generative framework that leverages GAN-guided diffusion to expose weaknesses in thermal detectors through controlled object and context level perturbations. Our approach learns class-conditional thermal priors via a GAN and uses diffusion to transform object appearances into ambiguous or low-saliency patterns. Simultaneously, a context module degrades scene conditions by simulating harsher night environments. Integrated with a detector-in-the-loop framework, our method generates realistic yet challenging inputs that reveal failure modes beyond traditional benchmarks. Experiments show significant performance degradation, while incorporating these scenarios into training improves robustness and generalization, highlighting the importance of semantic and context-aware evaluation in thermal perception.
Comments
Advisor: Sanjay Kumar Madria, madrias@mst.edu