Adversarial Robustness of Foundation Models for Intelligent Mechanical Systems: Threat Models, Benchmarks, and Defense Stacks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Foundation models increasingly operate across modalities (vision, language, audio, and vision–language) and are deployed in decision-critical pipelines with tool use and retrieval. This expands the adversarial surface: small perturbations to images or audio can flip predictions, carefully crafted text can induce unsafe actions, and cross-modal attacks can exploit representation alignment to produce consistent but wrong outputs. This paper reviews adversarial robustness of foundation models across modalities and proposes a unified benchmark-and-defense stack. We first formalize multimodal threat models (white-box/black-box, digital/physical, prompt-level/system-level) and show how attack objectives differ across classification, retrieval, captioning, and agentic planning. We then summarize benchmark families for robustness: standardized perturbation budgets in vision, imperceptible audio attacks, instruction-following adversarial prompts in language, and cross-modal attacks on vision–language alignment and retrieval. Finally, we present a practical defense stack combining (i) robust training and regularization, (ii) multimodal input sanitization and consistency checks, (iii) retrieval/verification and ensemble critics, and (iv) runtime guardrails for tool execution. We recommend reporting both utility and security metrics: clean accuracy, robust accuracy, attack success rate, confidence calibration, and worst-case safety violations under adaptive adversaries. The goal is to provide a deployment-oriented roadmap for measuring and improving robustness of multimodal foundation models.