Ablation experiments.

This study proposes a novel feature extraction framework that integrates reinforcement learning-guided steganographic encoding with an improved EfficientNetV2 backbone, specifically tailored for sustainable accounting and environmental auditing tasks. By embedding a domain-adaptive multi-branch attention mechanism and leveraging a lightweight residual policy network, the model is capable of capturing subtle patterns in noisy, imbalanced, and partially missing datasets. Experimental results on three real-world ESG-related (Environment, Society, Governance) accounting datasets demonstrate that the proposed method outperforms state-of-the-art models in terms of classification accuracy, robustness, and explainability. The model achieves an average AUC (Area Under the Curve) improvement of 4.7% and a 12.5% reduction in feature redundancy. Additionally, it exhibits superior performance in privacy-constrained scenarios through embedded steganographic masking. These findings underscore the framework’s potential for real-world deployment in regulatory auditing, automated compliance, and sustainable financial intelligence.