Mulkan Fadhli, Teuku Yuliar Arif, Ramzi Adriman, Roslidar
Machine Learning as a Service (MLaaS) has become a widely adopted paradigm for delivering predictive models through APIs, but it exposes these models to steal. Adversaries can replicate proprietary models using query-based, auxiliary-data, or data-free attacks, undermining intellectual property and privacy. Existing defenses such as Differential Privacy (DP), Adversarial Training, and Rate Limiting mitigate leakage but remain insufficient against adaptive and stealthy strategies. This paper proposes a conceptual defense framework that integrates DP with anomaly-aware detection to address model stealing. In the proposed design, incoming queries are simultaneously processed by a noise-injection mechanism and analyzed by an anomaly detection pipeline consisting of Autoencoders (AE) and Graph Neural Networks (GNNs). The AE captures statistical irregularities in individual queries, while the GNN models relational and temporal dependencies across sessions to detect stealthy or coordinated attack patterns. A Response Decision unit fuses these signals to either allow benign queries or block suspicious ones in real time, limiting information leakage and hindering substitute model training. As a proof of concept, preliminary experiments on mixed benign and adversarial query traffic show that both the AE’s reconstruction error and the GNN’s anomaly score clearly separate normal queries from attack traffic, achieving ROC AUC values above 0.98, supporting the approach’s feasibility. The contribution of this work is a unified architecture combining privacy-preserving perturbation with adaptive anomaly detection, targeting auxiliary traffic and evolving adversarial behaviors. Although still at a conceptual stage, the framework offers a blueprint for resilient MLaaS defenses against both high-frequency and low-and-slow model stealing strategies. ©2025 IEEE.
Doctoral School of Engineering Science, Universitas Syiah Kuala, Banda Aceh, Indonesia; Information Technology Depart., Universitas Islam Negeri Ar-Raniry Banda, Aceh, Indonesia; Electrical and Computer Engineering Department, Universitas Syiah Kuala, Banda Aceh, Indonesia
Research at a Glance
Register to unlock
Topics & SDG Alignment
Register to unlock
Collaboration
Register to unlock
Author Profile (Selected)
Register to unlock
References Overview
Register to unlock
Journal & Source
Register to unlock
Metadata & Integrity
Register to unlock