Adaptive Failure Mode Synthesis for Proactive Robustness Testing

While Le Coz et al. (2024) and Yigin & Çelik (2024) demonstrate using GANs to discover known failure conditions, this research proposes a fundamentally different approach: adaptive failure synthesis that generates novel failure modes beyond those seen in training data. The core innovation lies in training a "failure generator" that learns the boundaries of what constitutes a failure case, then deliberately explores just beyond these boundaries to create unseen edge cases. Unlike existing failure analysis that works retrospectively, our method would use reinforcement learning to actively search the latent space for maximally surprising yet plausible failure modes. This builds on the failure detection work but adds a proactive dimension - essentially creating a "red team" AI that thinks of ways to break systems before real users do. The approach could be particularly valuable for safety-critical applications like medical diagnosis systems (building on Ctrl-GenAug's medical focus) or autonomous vehicles, where discovering novel failure modes before deployment could save lives.

References:

1. Leveraging generative models to characterize the failure conditions of image classifiers. Adrien Le Coz, Stéphane Herbin, Faouzi Adjed (2024). AISafety@IJCAI.
2. A Prescriptive Model for Failure Analysis in Ship Machinery Monitoring Using Generative Adversarial Networks. Baris Yigin, Metin Çelik (2024). Journal of Marine Science and Engineering.
3. Ctrl-GenAug: Controllable Generative Augmentation for Medical Sequence Classification. Xinrui Zhou, Yuhao Huang, Haoran Dou, Shijing Chen, Ao Chang, Jia Liu, Weiran Long, Jian Zheng, Erjiao Xu, Jie Ren, Ruobing Huang, Jun Cheng, Wufeng Xue, Dong Ni (2024). arXiv.org.