Generative Backdoor Optimization and LLM-Based Recommendation for Robust Intelligent Systems

Abstract

Robust intelligent systems must address both external adversarial manipulation and internal data contamination. Universal transferable attacks show that vision-language models can be vulnerable to targeted adversarial examples, while clean-image backdoors demonstrate that malicious triggers can be optimized without obvious visual artifacts. These risks are particularly relevant for systems that combine multimodal models, personalized recommendation, and lifelong learning. LLM-based sequential recommendation introduces dynamic indexing mechanisms that improve user modeling but also expand the attack surface for poisoned data, biased inputs, and unreliable outputs. Biomedical evidence on remote cardiac injury provides an example of complex high-stakes reasoning where trustworthy learning is essential. Flood disaster relocation research adds a public-policy context in which intelligent systems may support decisions affecting households and communities. This literature cluster emphasizes adversarial robustness, dynamic recommendation, biomedical reasoning, and disaster-oriented decision support.