DeMas: An efficient method for malicious samples detection and mitigation in cloud-based systems

Abstract

Cloud services, particularly large-scale computing and data platforms, have become integral to enterprise operations, processing vast volumes of input data in real-time. However, these systems are increasingly vulnerable to adversarial actors who inject malicious data, thereby posing substantial security threats. Prevailing detection mechanisms often emphasize unintended class exclusions, which are inadequate in mitigating malicious attacks and are especially susceptible to class imbalance. To overcome these limitations, we introduce DeMas, a novel framework for the detection and mitigation of malicious samples. DeMas synergistically integrates adversarial perturbation with neighborhood averaging to robustly identify anomalous inputs. Furthermore, it employs a diffusion model, guided by a tractable probabilistic model, to remediate identified threats at the input level. This dual-stage approach transforms malicious samples into benign counterparts, thereby enhancing the security of downstream cloud-based models while preserving the usability of the data. Our empirical evaluation demonstrates that DeMas achieves a detection accuracy of 91.37% on a dataset of malicious samples, affirming its efficacy as a comprehensive defense strategy for secure and scalable cloud computing environments.