In response to the problem of insufficient adaptability of parameterized learning in local observation graph signal scenes in multi-task learning scenarios, a graph signal processing technique for multi-task scenarios is studied and introduced. In this technology, the research is based on the improved diffusion least mean square algorithm for clustering sampling, using different graph filters to train parameters and a clustering strategy to complete sampling. In addition, a multi-task power grid voltage graph signal detection technique with improved high pass filtering is proposed, and power grid data security detection is achieved through threshold setting and projection detection. In the analysis of graph signal processing, the research model has better steady-state learning performance. For example, in the transient mean square deviation test of multitasking scenarios, its convergence value is 31.25 dB, which is better than similar models. In the single-task transient mean square deviation test, the research model is closer to the reference standard, with a convergence value of 34.25 dB and better learning ability. In power application analysis, the research model performs better in graph signal detection, such as in voltage amplitude scenarios where the maximum detection probability is 100%. In the voltage angle scenario, when the voltage angle is higher than ±2°, the detection probability of the research model is higher than 92.25%. It can be seen that the technology proposed by the research has excellent application effects. This study will provide technical support for the improvement of graph signal technology and data security.