Distributed secondary control has been proposed to maintain frequency/voltage synchronization and power sharing for distributed energy sources in AC microgrids (MGs). The cyber layer is susceptible to time delays and cyber failures and thus, a distributed resilient secondary control should be investigated. This paper proposes a distributed multi-scale attention and predictor-based control (DMAPC) strategy to address false data injection attacks and packet loss failures with time delays. The multi-scale attention mechanism enables the system to selectively focus on neighbors ’ states with higher confidence evaluated in different time scales, while the data-driven predictor compensates for lost neighbors’ states in the nonlinear controller. The DMAPC does not impose strict limitations on the number of false communication links or upper bound for false data. Besides, the DMAPC is formulated as an uncertain system with time delays and is proven to be uniformly ultimately bounded. Extensive experiments on a hardware-in-the-loop MG testbed have validated the effectiveness of DMAPC, which successfully relaxes restrictions on cyber failures compared to existing strategies.