Adversarial Robustness and SLA Security in Cost-Aware Auto-Scaling: Mitigating Policy-Poisoning Attacks on DRL Controllers in Multi-Cloud Deployments

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

Deep reinforcement learning (DRL) controllers have emerged as promising solutions for cost-aware auto-scaling in multi-cloud environments, yet their vulnerability to adversarial manipulation remains critically under-examined. This study addresses the research gap concerning policy-poisoning attacks targeting DRL-based auto-scaling controllers, wherein adversaries manipulate reward signals during training to induce economically devastating scaling decisions. We propose a guardrailed ensemble framework that integrates anomaly detection mechanisms with robust policy validation protocols. Through extensive simulation using real-world workload traces across three major cloud providers, our proposed defense achieves 89.4% detection accuracy for poisoned policy states while maintaining 96.2% of baseline cost efficiency. Statistical analysis reveals that guardrailed controllers exhibit 73% reduction in SLA violation severity compared to unguarded DRL baselines under adversarial conditions. The findings demonstrate that policy-poisoning attacks can increase operational costs by 42-67% through induced over-provisioning, while our guardrail framework contains such attacks to within 8.3% of optimal cost. This research contributes a replicable security architecture for production DRL orchestration systems, offering practitioners a validated approach to safeguarding autonomous scaling infrastructure against emerging adversarial threats in multi-cloud deployments.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.