Cloud-Native Deep Learning Architectures For Secure Generative AI Deployment In Enterprise Workflow Platforms

Abstract

Generative artificial intelligence (GenAI) is a rapidly developing technology with potential for multiple applications, yet it is complex, resource-intensive, and prone to risks. Deployment of GenAI in enterprise workflow platforms requires approaches that enable secure operation while maintaining availability, reliability, and quality. A synthesis of cloud-native architectural patterns with contemporary risk frameworks provides insights into essential security aspects for GenAI. Findings indicate that careful consideration of the safeguards available for prompt injection mitigation, model inversion protection, and data privacy when developing GenAI within a service-mesh architecture can reduce the likelihood of future attack success or damage.

Cloud-native generative artificial intelligence (GenAI) deployment in enterprise workflow platforms is increasingly common, especially for support documentation creation. However, safeguarding the system against attacks that target the availability, reliability, or data privacy of the service remains challenging. Leveraging cloud-native patterns of scalability, resilience, composability, portability, and observability can guide security-enhancing measures. Mapping the security-by-design concept to GenAI services within a service-mesh architecture identifies a range of security controls rooted in established identity and access management principles, the concept of security through obscurity, the defense-in-depth principle, and the auditing of logs and monitoring alerts.