To address the challenges of high collision risk and insufficient decision-making robustness for Connected and Automated Vehicles (CAVs) during mandatory lane changes in expressway weaving areas, this paper proposes a decision-making model that integrates a composite risk field with an adversarial game mechanism. Firstly, real-time quantified dynamic risk values are mapped onto the state space and multi-objective reward function of the Dueling Deep Q-Network (Dueling DQN) algorithm, endowing the agent with explicit risk perception capabilities. Secondly, an adversarial agent based on the Proximal Policy Optimization (PPO) algorithm is designed to proactively generate safety-critical scenarios via a zero-sum game framework. Furthermore, an "Ego-Adversary" alternating evolutionary training strategy is adopted to expose decision-making vulnerabilities. Simulation results demonstrate that the Dueling DQN outperforms baseline algorithms across various traffic flow densities, maintaining a lane-changing success rate of 90.0% at medium density (P=0.5). Optimal robustness is achieved at an adversarial intensity of 0.4, where the collision rate is reduced to below 5%, and lateral collisions are effectively suppressed. Following multi-round adversarial training, performance metrics show significant convergence. Spatiotemporal trajectory analysis confirms that the proposed strategy effectively inhibits the generation of traffic shockwaves, facilitating a transition from disordered congestion to a stable synchronized flow state and reaching a Nash Equilibrium. Ultimately, a highly robust decision-making strategy is achieved that balances safety, efficiency, and comfort.