Integrity Verification Framework for User-Subscribed AI Models on the Edge Platform

• Published in: IEEE International Conference on Edge Computing (Online) pp. 54 - 63
• Main Authors: Oyeniyi, Oluwafeyisayo, Roy, Ayan, Sen, Amartya
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.07.2025
• Subjects: Artificial intelligence; Blockchain; Computational modeling; Consensus algorithm; Data models; Edge computing; Edge Intelligence; Fault tolerant systems; Nearest neighbor methods; Servers; Throughput; Time factors
• ISSN: 2767-9918
• DOI: 10.1109/EDGE67623.2025.00015

The incorporation of Artificial Intelligence with edge computing platforms is defined as edge intelligence (EI). The rising popularity of EI necessitates the need to ensure the integrity of the AI models used for EI. However, the resource-constrained, decentralized, and distributed nature of edge computing poses significant challenges in verifying the integrity of AI model parameters, as traditional verification mechanisms cannot be implemented. This makes the AI models vulnerable to data poisoning, collusion attacks, and data inference attacks, which undermines their trustworthiness. Without effective verification mechanisms, AI models become vulnerable to data inference and data poisoning, undermining the reliability and trustworthiness of AI-driven applications used in EI-based applications. To address these challenges, this paper introduces the architectural concept of AI-as-a-service for Edge platforms and then proposes a framework developed using a Blockchain-based (Proof of Authority) consensus algorithm to verify the integrity of on-demand AI models that will be subscribed by the users. These on-demand AI models further reduce the risk of data breaches, as the data of the user stays local, provides faster response time, and offers flexibility in scaling AI resources. Simulations have been carried out using five different datasets, executing six different machine learning-based AI models like SVM, LR, and KNN, to name a few. The effectiveness of the proposed framework is highlighted by the fact that the proposed framework is resilient to false information inference attacks when less than 50% of the edge servers are compromised in a network. The efficacy of the proposed framework was further evaluated by analyzing fault tolerance (84.7% success rate) and throughput performance.