Abstract:In the era of big data, the storage and retrieval of massive data have become a critical bottleneck for system performance. Although learned indexes hold widespread application in large-scale data management through their high query efficiency and low space overhead, their performance is significantly degraded under skewed local data distribution, especially in the presence of poisoning attacks. Traditional learned indexes rely on an in-place insertion strategy with fixed thresholds. However, to maintain local sorted order, random insertion may trigger costly retraining and data movement operations, which consume substantial computational resources and result in unstable system responsiveness and reduced throughput. To this end, this study proposes a buffered point smoothing weighted linear regression insertion (BPS-WLRI) mechanism. Firstly, this method introduces a weighted linear regression model to finely fit the data distribution, thereby effectively reducing the index tree depth and retraining frequency. Meanwhile, a hash-bucket buffer structure is employed to optimize the in-place insertion process, effectively mitigating the large-scale data movements caused by strict maintenance of sorted order. Additionally, a buffer-point smoothing strategy is designed to ensure that during buffer merging, only local structures are incrementally updated, thereby avoiding the high overhead of global restructuring. Experimental results demonstrate that under extreme poisoning attack loads, BPS-WLRI achieves throughput improvements of 1.81 times, 4.32 times, 2.22 times, 2.06 times, 1.41 times, and 1.39 times over ALEX, PGM, LIPP, DILI, SALI, and AULID respectively. These results confirm the efficiency and stability of the proposed algorithm under dynamic data loads and malicious attack scenarios.