Public Integrity Verification for Run-time Model in Batches without a Local Server

Current integrity verification schemes for intelligent model primarily rely on static or dynamic monitoring methods that target a single server. However, in practical scenarios, developers opt to deploy intelligent models across multiple remote servers provided by server providers to achieve lower latency and enhanced resilience against risks. Consequently, there is a pressing need for a dynamic monitoring approach that can effectively oversee multiple remote servers, addressing the concern of both developers and users. In this paper, we propose a public integrity solution designed for run-time deep models in batches without a local server. The core of this method lies in the generation of sensitive samples, followed by the aggregation of their model prediction results to create the model fingerprint. The experimental results shows that the method proposed in this paper can effectively detect whether the model has been tampered with as long as the number of compromised servers does not exceed half of the total.