Abstract

Autonomous agents operated across distributed and dynamic environments faced increasing exposure to security threats that disrupted decision integrity and system reliability. Prior studies indicated that adversarial interference, policy manipulation, and input perturbation affected agent behaviour, although systematic protection remained limited. The study addressed this gap by examining structured defence mechanisms for secure autonomous decision processing. The problem focused on inconsistent threat recognition and weak resilience against adversarial manipulation within multi- agent environments. Existing approaches lacked unified modelling for threat isolation and response stability under uncertain conditions. To address this issue, the study proposed an Adaptive Adversarial Isolation Framework (AAIF), which integrated multi-layer threat observation, anomaly scoring, and isolation-based response logic. The framework did not rely on static rule definitions and instead maintained adaptive separation between normal and malicious behavioural patterns. The method followed a structured analytical design where simulated autonomous agent environments underwent controlled adversarial injection scenarios. The AAIF model did analyse behavioural deviations through probabilistic mapping and comparative state evaluation across operational cycles. Performance evaluation was conducted using stability metrics, detection consistency measures, and response latency indicators. Results indicated that AAIF achieved improved threat separation consistency and maintained stable operational outputs under adversarial pressure conditions. The system reduced incorrect behavioural propagation across connected agents and strengthened decision robustness under uncertain inputs. Comparative analysis showed that AAIF sustained higher detection reliability across varied attack scenarios when compared with baseline defensive configurations.

Authors

J. Jasmine¹, Thinanath Ravichandran²
Karpagam College of Engineering, India¹, Monash University, Australia²

Keywords

Autonomous Agents, Adversarial Security, Threat Detection, Adaptive Framework, Multi-Agent Systems