Flocking constitutes a typical collective motion behavior observed within biological communities, finding extensive applications in fields such as drone swarms. Multi-agent deep reinforcement learning offers a novel approach to address the difficulties of parameter tuning, but still confronts issues of non-stationarity and poor scalability constraints. Starlings can emerge complex collective intelligence through simple individual behaviors facilitated by the nature of nearest-neighbor interaction. This paper proposes a starling-inspired distributed deep reinforcement learning approach for multi-agent collective flocking. Specifically, the integration of k-nearest neighbor interaction and attention mechanisms mimics the interaction characteristics among starlings, enhancing the scalability and the stability of proposed model. Individual reward functions are designed inspired by Boids rules, with 3 small-scale individual network architectures are proposed, to simulate the emergence of complex collective behaviors. Comparative experiments validate the effectiveness and scalability of the proposed strategy.