Android is a popular target for malware attacks due to its open-source and open nature. However, machine learning-based malware detection systems are vulnerable to adversarial sample attacks. The generation of adversarial sample primarily focuses on methods such as adding API perturbations, redundant code, and permissions. Nevertheless, these attack methods suffer from improper execution and corruption of the original malicious functions. In this paper, we investigate malware adversarial techniques, design an algorithm (DERGB algorithm) for generating adversarial samples targeting Android malware, based on RGB images, and utilize the Differential Evolution (DE) algorithm to identify the pixels in the image that impact the classification results for perturbation. Our algorithm aims to minimize the number of pixels requiring perturbations. Additionally, we leverage the header information class.dex file features, which contain the data area offsets, to constrain the range of pixel modifications in the algorithm for adversarial sample generation. This ensures that the malicious functionality and executability of the adversarial sample software remain uncompromised. The Experimental results demonstrate that our algorithm incurs a decrease in the accuracy of the detection model by approximately 34.8 \(\%\) when subjected to an attack by a single pixel point, and about 77.14 \(\%\) when subjected to five-pixel point attacks. These findings demonstrate that our algorithm achieves a comparable attack effect to other anti-sample attack algorithms, even with minimal perturbations.