This paper addresses bandwidth limitations resulting from Denial-of-Service (DoS) attacks on Artificial Intelligence of Things (AIOT) systems, with a specific focus on adverse network conditions. First, to mitigate the impact of DoS attacks on system bandwidth, a novel model predictive control combined with a dynamic time-varying quantization interval adjustment technique is designed for the encoder–decoder architecture of AIOT systems. Second, the network state is modeled to represent a Markov chain under suboptimal network conditions. Furthermore, to guarantee the stability of AIOT systems under random packet loss, a Kalman filter algorithm is applied to precisely estimate the system state. By leveraging the Lyapunov stability theory, the maximum tolerable probability of random packet loss is determined, thereby enhancing the system’s resilient operation. Simulation results validate the effectiveness of the proposed method in dealing with DoS attacks and adverse network conditions.