Middleware Architecture for the Management and Mitigation of OWASP ML05: Model Theft in IoT Machine Learning Networks

The increasing integration of machine learning (ML) models into Internet of Things (IoT) applications has led to notable advancements in automation and decision-making. However, these models are vulnerable to modern attack vectors recognized by the OWASP Top 10 for Large Language Model Applications, specifically ML05: Model Theft, where adversaries gain unauthorized access to model parameters and training data, compromising intellectual property and sensitive information. Such threats are particularly concerning in IoT environments due to their distributed nature and resource limitations. This paper proposes a middleware architecture for the management and mitigation of model theft risks by incorporating encryption, access control, obfuscation, watermarking, continuous monitoring, and service assurance programmability. By strengthening the security management framework of ML models deployed in IoT, the proposed architecture aims to protect against theft, ensure data confidentiality, and maintain network resilience. The approach includes detailed mathematical models and an evaluation of existing security measures, demonstrating the architecture's effectiveness in diverse IoT deployments, such as telemedicine and smart cities.