Location
Havener Center, Miner Lounge / Wiese Atrium, 9:30am-11:30am
Start Date
4-2-2026 9:30 AM
End Date
4-2-2026 11:30 AM
Presentation Date
April 2, 2026; 9:30am-11:30am
Description
Incremental learners deployed on streaming data must remain robust to evolving adversarial perturbations, yet most adversarial-robustness studies assume offline multi-epoch training with repeated access to historical data. We investigate adversarial robustness in Fuzzy ARTMAP, a prototype-based Adaptive Resonance Theory model that supports single-pass learning without replay. We propose WB-Softmax, a differentiable relaxation that aggregates category-level activations into class-level scores for gradient-based attacks. WB-Softmax PGD achieves 89–100% attack success on vanilla models, exceeding transfer and query-based baselines. We then study adversarial training under true streaming constraints by comparing offline versus online adversarial example generation and standard versus selective updates. Offline adversarial training consistently collapses robustness, whereas online training is dataset-dependent. We introduce progressive two-stage selective training that achieves the best overall robustness on all three datasets. Finally, we monitor incremental cluster validity indices to diagnose separation collapse and design a separation-aware absorption rule that improves high-epsilon robustness.
Biography
Shane is a 3rd year PhD student and Kummer I&E Doctoral Fellow in the Computer Science Department, advised by Dr. Don Wunsch in the Applied Computational Intelligence Laboratory. His research focuses on adversarial robustness and interpretability in incremental learning systems, particularly Adaptive Resonance Theory networks, with broader interests in AI safety and policy. Prior to his doctoral studies, Shane gradauted from S&T with a B.S. in Computer Science and gained industry experience through internships at Ford and Howmet Aerospace. Outside of research, he enjoys soccer, chess, and hiking.
Meeting Name
2026 - Miners Solving for Tomorrow Research Conference
Department(s)
Electrical and Computer Engineering
Second Department
Computer Science
Document Type
Poster
Document Version
Final Version
File Type
event
Language(s)
English
Rights
© 2026 The Authors, All rights reserved
Robustness of Fuzzy ARTMAP to Adversarial Attacks and Progressive Adversarial Training for Streaming Learning
Havener Center, Miner Lounge / Wiese Atrium, 9:30am-11:30am
Incremental learners deployed on streaming data must remain robust to evolving adversarial perturbations, yet most adversarial-robustness studies assume offline multi-epoch training with repeated access to historical data. We investigate adversarial robustness in Fuzzy ARTMAP, a prototype-based Adaptive Resonance Theory model that supports single-pass learning without replay. We propose WB-Softmax, a differentiable relaxation that aggregates category-level activations into class-level scores for gradient-based attacks. WB-Softmax PGD achieves 89–100% attack success on vanilla models, exceeding transfer and query-based baselines. We then study adversarial training under true streaming constraints by comparing offline versus online adversarial example generation and standard versus selective updates. Offline adversarial training consistently collapses robustness, whereas online training is dataset-dependent. We introduce progressive two-stage selective training that achieves the best overall robustness on all three datasets. Finally, we monitor incremental cluster validity indices to diagnose separation collapse and design a separation-aware absorption rule that improves high-epsilon robustness.
Comments
Advisor: Donald C. Wunsch, dwunsch@mst.edu