AbstractWe propose a novel dual-stream neural architecture for secure end-to-end communication that integrates self-supervised adversarial detection with game-theoretic defense strategies. The increasing sophistication of adversarial attacks necessitates adaptive security mechanisms that can dynamically respond to evolving threats while maintaining computational efficiency. Our framework addresses this challenge by combining a transformer-based contrastive learning module for real-time perturbation detection with a zero-sum Markov game formulation to optimize defense policies under Nash equilibrium. The self-supervised stream extracts invariant features from communication signals and computes threat probabilities, while the game-theoretic stream dynamically adjusts encryption parameters and defense actions based on adversarial behavior. The two streams are jointly optimized to achieve provable robustness against both known and unknown attack patterns. Furthermore, the system interfaces seamlessly with conventional intrusion detection and encryption modules, replacing rule-based heuristics with data-driven adaptive strategies. Experimental validation demonstrates significant improvements in detection accuracy and resilience compared to existing methods. The key innovation lies in the unification of self-supervision and game theory within a single end-to-end trainable framework, enabling scalable and adaptive security for modern communication networks. This approach not only enhances real-time threat mitigation but also provides theoretical guarantees on defense optimality under adversarial dynamics.